[libs] Skip AdvSimdEncode on Mono #96829
Conversation
dotnet-issue-labeler
bot
added
the
needs-area-label
ghost
assigned 
|
/azp run runtime-extra-platforms |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
Shouldn't the #if MONO be in the caller ? |
| @@ -493,6 +493,9 @@ private static unsafe void Avx2Encode(ref byte* srcBytes, ref byte* destBytes, b | |||
| [CompExactlyDependsOn(typeof(AdvSimd.Arm64))] | |||
| private static unsafe void AdvSimdEncode(ref byte* srcBytes, ref byte* destBytes, byte* srcEnd, int sourceLength, int destLength, byte* srcStart, byte* destStart) | |||
| { | |||
| #if MONO | |||
There was a problem hiding this comment.
It is better to do this at the callsite of AdvSimdEncod.
There was a problem hiding this comment.
I can add an extra guard to the current caller of AdvSimdEncode. I think the current callsite would just fall through to the next block of logic, so I didn't add it in the first place.
I think we should still have the block on AdvSimdEncode in case more invocations are added, because the PNSE is hit inside of AdvSimdEncode.
There was a problem hiding this comment.
hhm, wondering why this should be at the callsite rather than just inside AdvSimdEncode? Also can you add a comment about why this is under a #if !MONO pointing to the #96828 (comment)?
In caller, you could just have without #else
#if MONO
return;
#endifThere was a problem hiding this comment.
Because at the callsite, there is a check to see if AdvSimd is supported to decide to call AdvSimdEncode or not. So I feel logically, it makes more sense to do it there. Functional wise, either way should be fine.
There was a problem hiding this comment.
pointing to the #96828 (comment)?
point it to #93081
there is a check to see if AdvSimd is supported
Right, but that is true even for Mono and should be ok. At least if there are more callers, we won't have to worry about adding the #if everywhere.
There was a problem hiding this comment.
Updated to guard the entire method
I believe that Mono reporting IsSupported but functions not being supported might be attributed to #84510 not accounting for Mono. The volume of mono tests is far too time consuming + non-zero flakey tests to run on all PRs by default. Hence a runtime-extra-platforms lane was introduced, which was intended to be ran for PRs that make changes affecting Mono.
There is a smoke test being ran by default, but that is just System.Runtime.Tests.csproj. We could increase the number of suites that are ran as smoke tests, but that is another question of what other suites should be added that reliably dont flake. It looks like there are HWIntrinsics runtime tests, but guessing it doesn't run for Mono?
There was a problem hiding this comment.
I don't think mono specific pipelines are run today if the changes are not related to mono. runtime-extra-platforms would be a good remainder to run for such cases. Regarding, IsSupported, what is a good option here? Although logically correct to have AdvSimd.IsSupported == false until those APIs are not implemented in mono, but that would mean that mono won't have any of the other AdvSimd features too.
There was a problem hiding this comment.
BCL change should trigger mono CI lanes which run library tests to run. If not, it is an issue worth looking into.
There was a problem hiding this comment.
Since these are new API's, I doubt the customer would use them right away. So I would prefer leaving AdvSimd.IsSupported == true as it is now and implement the new API's soon.
There was a problem hiding this comment.
re: tests, I think this is a case where we didn't have adequate coverage in our smoke tests. The good thing is that we discovered it right away via extra-platforms and we can take action.
With us running a reduced set of tests for iOS devices (full aot LLVM), this is going to happen. As Mitch was getting at, my suggestion would be to add another suite to the smoke tests that we run so that there's a reasonable chance we would fail on something like this in the future.
am11
added
area-System.Buffers
and removed
needs-area-label
|
Tagging subscribers to this area: @dotnet/area-system-buffers Issue DetailsWorks around #96828 Not sure if the problematic code affects all of mono arm64 configurations or just ios/tvos arm64.
|
mdh1418
force-pushed
the
skip_AdvSimdEncode_mono
branch
from
aa87122
to
b9556d0
Compare
|
/azp run runtime-extra-platforms |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run runtime-extra-platforms |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
Runtime failures are known, and runtime-extra-platforms failures are known/not related. |
|
Given that the latest commits were a matter of how AdvSimdEncode was excluded on Mono, I believe the test results wouldn't change from the previous run, and this PR is essentially excluding #95513 on Mono, so once @fanyang-mono @kunalspathak @tannergooding does excluding all of the code introduced by #95513 look good? |
This reverts commit 1a76e37.
* [libs] Skip AdvSimdEncode on Mono * Add guard around existing AdvSimdEncode callsite to exclude on Mono * Add issue to guard * Guard AdvSimdEncode logic for not Mono * Make AdvSimdEncode inaccessible on Mono * Update issue comment
…#96944) * Revert "[libs] Skip AdvSimdEncode on Mono (dotnet#96829)" This reverts commit 1a76e37. * Revert "Use multi-reg load/store for EncodeToUtf8 (dotnet#95513)" This reverts commit fdb03ca. * Wrap load/store vector APIs in '#if false' * Disable load/store vector tests * remove the trailing space
Works around #96828
Not sure if the problematic code affects all of mono arm64 configurations or just ios/tvos arm64.