gf: accelerate adding backedges back to old performance #48966
Conversation
Since it does not compute and branch on typeof, in my measurements, this implementation is about 10x faster!
vtjnash
added
compiler:latency
|
Did this regress in #46010? If so, maybe I should think about adding a benchmark. It doesn't seem easy, though, because it's not a step that's easy to isolate. |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
|
As you can see in the report, it actually had significantly regressed a benchmark (many_local_vars, née quadratic), but we hadn't done anything about it yet dd375e1#commitcomment-83236324 |
|
Ah, that does ring a bell. Thanks for explaining, and especially for fixing it! |
|
Nice improvements! Do you think it would be worthwhile to try implementing a similar optimization at the Julia level? Specifically, the inlining code often adds duplicated backedges, but we could use IdSet there to prevent the duplication. |
Since it does not compute and branch on typeof, in my measurements, this implementation is up to 10x faster! (cherry picked from commit 369660f)
Since it does not compute and branch on typeof, in my measurements, this implementation is about 10x faster!
@nanosoldier
runbenchmarks("inference", vs=":master")