Uh oh!
There was an error while loading. Please reload this page.
ARM/ARM64: Optimize virtual call stub for R2R and JIT - #36817
Conversation
kunalspathak
commented
May 23, 2020
@dotnet/jit-contrib , @jeffschwMSFT , @richlander , @davidwrighton , @jkotas |
kunalspathak
commented
May 26, 2020
@dotnet/runtime-infrastructure - I started the runs few days back and they are still showing as running. Is there a known issue or do I need to cancel/rerun the jobs? |
safern
commented
May 26, 2020
Yeah, that is a GH bug that is already reported. If you go into Azure DevOps, the build shows as finished and there are two failing legs: https://dev.azure.com/dnceng/public/_build/results?buildId=657536&view=results |
kunalspathak
commented
May 27, 2020
@BruceForstall , @CarolEidt - can you please take a preliminary look? I will see why Linux arm64 Release failed. The windows leg passed. |
safern
commented
May 27, 2020
@kunalspathak I believe you unfortunately ran the rolling build on a state where it was broken. The linux_musl_arm64 issue you're seeing was fixed and such as the browser_wasm issue. So I would recommend rebasing on top of master and queueing another build. |
This reverts commit 441d866609b406c57ed586a79f5ed27e99c31194.
This reverts commit 0183e76e83d5ba1057b95ba37204cc525e295bf5.
kunalspathak
commented
May 27, 2020
Thanks @safern for saving my day. :) |
safern
commented
May 27, 2020
@kunalspathak FYI... we will soon have linux_arm, linux_arm64 and linux_musl_arm64 libraries test runs using a checked coreclr when coreclr is changed: #36910 |
Does it impact my existing runs? Or are you suggesting that I might have to do another run when #36910 is merged? |
safern
commented
May 27, 2020
No it doesn't impact any of your runs. It was just an FYI for future PRs, that you'll get this sort of testing on your PR by default if you changed anything under coreclr. |
kunalspathak
commented
May 27, 2020
Ah, so I won't have to manually trigger |
safern
commented
May 27, 2020
IIRC, not yet. I'll double check. |
CarolEidt
left a comment
There was a problem hiding this comment.
I'd just like to see a description of what's being done (the why and how) for clarity.
Otherwise LGTM.
Uh oh!
There was an error while loading. Please reload this page.
kunalspathak
commented
May 27, 2020
@BruceForstall - As pointed out in #35108 (comment) there were 615732 redundant adrp/add pairs through out framework libraries. With work done in #35675 and in this PR, I just verified that there are no redundant adrp/add pairs anymore. |
Optimize the calls to virtual stub by eliminating the redundant load of address. This is similar to #35675 but optimizes virtual call stubs which impacts both crossgen/JIT.
Fixes: #36700
CrossGen improvements:
JIT improvements: