JIT: clear loop iter info after rebuilding loops for opt repeat #95879
Conversation
This info gets computed when finding loops but is not used by the general optimizer (it only cares about loop structure, not iteration details). If we leave it built it can go stale. Fixes dotnet#95843.
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThis info gets computed when finding loops but is not used by the general optimizer (it only cares about loop structure, not iteration details). If we leave it built it can go stale. Fixes #95843.
|
|
The issue was exposed because #95705 changed how the stress mode "random" CSEs work, and so the initial round of CSE did not CSE an ARR_LEN that we normally always CSE. The subsequent loop rebuild then set the LP flag based on this, and then the second round of CSE did the ARR_LEN CSE, hence the flag and IR diverged. @BruceForstall PTAL Opt repeat issue only, no diffs expected. |
|
This is going to conflict with #95854... But feel free to merge this and I'll fix it up tomorrow (I am missing some more handling for |
There was a problem hiding this comment.
LGTM.
I should really try to get #94250 finished -- I'd fixed this a month or so back...
|
Not sure why build analysis is not clearing this, failure is #95875 and this PR is logged there. |
This info gets computed when finding loops but is not used by the general optimizer (it only cares about loop structure, not iteration details). If we leave it built it can go stale.
Fixes #95843.