perf(es/minifier): Make DCE analyzer parallel #9865
Conversation
|
kdy1
changed the title
kdy1
force-pushed
the
dce-par
branch
6 times, most recently
from
d71bf03 to
593d5e9
Compare
![kodiakhq[bot]](https://avatars.githubusercontent.com/in/29196?s=60&v=4){kind=small}
| @@ -1111,6 +1203,87 @@ impl VisitMut for TreeShaker { | |||
| } | |||
| } | |||
|
|
|||
| fn merge_data(data: Arc<ThreadLocal<RefCell<Data>>>) -> Data { | |||
There was a problem hiding this comment.
Actually, this function is slow and I want to optimize this function further.
| } | ||
| }; | ||
| } | ||
| } | ||
|
|
||
| /// Traverse the graph and subtract usages from `used_names`. | ||
| fn subtract_cycles(&mut self) { |
There was a problem hiding this comment.
But this function call is necessary.
|
I tried profiling, but it seems to have almost no effect on current benchmark suite, which only contains libraries. DCE is impactful for real-world apps, though. |
CodSpeed Performance ReportMerging #9865 will not alter performanceComparing Summary
|
PRMain |
Description:
The analyzer of the DCE pass can be parallelized in a similar way to mark-sweep GC. It currently uses
&mut petgraph::DiGraphMapto find reachable bindings, but we can split the graph creation into multiple threads.Also, I decided to amort allocations because
swc_parallel::joinis called an enormous amount of time, meaning that it will be slower if I allocate one hashmap from each closure. I'll use thread_local crate to reuse the hashmap and collect all of them after the traversal is done.