perf: optimize trace identifiers #5680
Conversation
DaniPopes
changed the title
| impl LocalTraceIdentifier { | ||
| pub fn new(known_contracts: &ContractsByArtifact) -> Self { | ||
| Self { | ||
| local_contracts: known_contracts | ||
| .iter() | ||
| .map(|(id, (abi, runtime_code))| (runtime_code.clone(), (id.clone(), abi.clone()))) | ||
| .collect(), | ||
| } | ||
| impl<'a> LocalTraceIdentifier<'a> { | ||
| pub fn new(known_contracts: &'a ContractsByArtifact) -> Self { | ||
| Self { known_contracts } |
There was a problem hiding this comment.
Remove useless re-allocations, we only need to be able to iterate over the 3 values (id, abi, code)
| TraceKind::Setup => { | ||
| (verbosity == 4 && result.status.is_failure()) || verbosity >= 5 | ||
| } | ||
| TraceKind::Deployment => false, |
There was a problem hiding this comment.
Make this exhaustive and a bit clearer
crates/forge/bin/cmd/test/mod.rs
Outdated
| if should_include || gas_reporting { | ||
| decoder.identify(trace, &mut local_identifier); | ||
| decoder.identify(trace, &mut etherscan_identifier); | ||
| decoder.decode(trace).await; |
There was a problem hiding this comment.
This is the main change, moved the identify calls from above the match to inside the if block
There was a problem hiding this comment.
this change is really weird, it speeds up significantly when there are a lot of traces, but not when there are little? I don't really know or have the resources to investigate, so I'll remove it from this PR
|
Needs a fix, i would've made it a draft if i knew how to on mobile |
Evalir
changed the title
|
@DaniPopes gotcha, reopened and marked it as |
DaniPopes
changed the title
DaniPopes
changed the title
| @@ -79,7 +79,7 @@ pub(crate) fn decode_cheatcode_inputs( | |||
| "serializeBytes32" | | |||
| "serializeString" | | |||
| "serializeBytes" => { | |||
| if verbosity == 5 { | |||
| if verbosity >= 5 { | |||
Motivation
Identifying and decoding traces contributes the majority (usually >40%) of the time spent in running tests. We can avoid most of this by just using the existing check flags to skip identifying only when needed.My assumptions were wrong. Generally, we decode traces on the main thread in parallel to the threads executing tests. This is a problem when there are a lot of them, and are sent batched (like for invariants). This causes the main thread to block for a while just decoding traces, which may also block after all tests are done.
This PR cleans up some tracing code, but the issue mentioned above will have to be fixed in another PR.
Solution