[NativeAOT] Avoid redundant unwind info lookup during stackwalks #87419
Conversation
Unwind info was looked up twice on non-Windows: Once during initial frame inspection and second during the actual unwind. Cache the lookup results from the initial frame inspection and use them from the actual unwind later, same as it is done on Windows.
|
Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas Issue DetailsUnwind info was looked up twice on non-Windows: Once during initial frame inspection and second during the actual unwind. Cache the lookup results from the initial frame inspection and use them from the actual unwind later, same as it is done on Windows.
|
|
Todo API w/ Workstation GC: Before this change: | Requests/sec | 72,812 | With this change: | Requests/sec | 76,727 | |
|
This is follow up on #86690 (comment) . I have more changes on top this one to make stack unwinding faster on Unix. |
|
Related: #75807 |
|
/azp run runtime-extra-platforms |
|
Azure Pipelines successfully started running 1 pipeline(s). |
I think the code makes assumption that fetching this info is a relatively cheap operation - like indexing in a table, but it is more complex than that. And there is space in the MethodInfo to keep the info. Nice change! |
ghost
locked as resolved and limited conversation to collaborators
3 participants
Unwind info was looked up twice on non-Windows: Once during initial frame inspection and second during the actual unwind. Cache the lookup results from the initial frame inspection and use them from the actual unwind later, same as it is done on Windows.