[release/8.0-rc2] Don't generate AddMask as it requires more explicit consideration of semantics#92308
Conversation
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsBackport of #92282 to release/8.0 /cc @tannergooding Customer ImpactTestingRiskIMPORTANT: If this backport is for a servicing release, please verify that:
|
|
@tannergooding is it your intention to target RC2? There's still time. The hard due date for merging is on Sept 24th. If yes, then you need to retarget your PR to |
|
This is a JIT side issue, so it likely needs determination from @JulieLeeMSFT or @jakobbotsch Given it's a correctness issue, it might be better to target RC2 |
|
Agreed, let's target rc2 for this one since it's customer reported. |
jeffschwMSFT
left a comment
There was a problem hiding this comment.
please get a code review
|
Approved by Tactics via email. The RC2 branch had a generalized failure for which we just merged a fix, so I updated this PR to ensure the CI is clean. |
Backport of #92282 to release/8.0-rc2
/cc @tannergooding
Customer Impact
The customer will experience incorrect codegen that results in the code doing something different than they intended when summing
Vector512<T>s created via certain patterns withoperator+. Customer reported in #92261.Testing
Manual validation that the reported customer scenario produces the expected codegen.
Risk
Low. This is simply removing an optimization that shouldn't have been included in the first place as it was incomplete. Additional work was required to ensure that the
kaddcode generation worked as expected for adding two masks together.