Reduce the number of hwintrinsic tests by removing no longer interesting scenarios #85008
Conversation
|
Tagging subscribers to this area: @dotnet/area-system-runtime-intrinsics Issue DetailsWe have a considerable number of intrinsics and therefore total tests that exist, however since the intrinsics were first introduced there have been many improvements to the general infrastructure such that many of the scenarios we were originally testing are no longer interesting. In particular, we previously had many different types of loads being tested to validate that containment and other specialized logic around varying addressing modes were correctly handled in the face of the customer Today, however, such loads are normalized to standard As such, this removes those scenarios from the core test templates so we can decrease the total run time for all hwintrinsic tests.
|
|
There are more cases of this exact scenario we could remove in the other templates, but the majority of them are covered by the 11 I've updated here. I'll log an item tracking us cleaning up the same in the other templates. There are likely also some other now uninteresting scenarios we could further remove, but those are less obvious and may need more data to justify their removal. |
|
CC. @BruceForstall, @dotnet/avx512-contrib, @dotnet/jit-contrib As per the top post, this is just a reduction in the hwintrinsic tests. With the addition of AVX-512 we were encroaching up on the timeouts again and as described above, these scenarios are no longer significantly interesting to test given changes we've made to the JIT over the past 6-8 months so can be safely removed. |
ghost
locked as resolved and limited conversation to collaborators
We have a considerable number of intrinsics and therefore total tests that exist, however since the intrinsics were first introduced there have been many improvements to the general infrastructure such that many of the scenarios we were originally testing are no longer interesting.
In particular, we previously had many different types of loads being tested to validate that containment and other specialized logic around varying addressing modes were correctly handled in the face of the customer
Loadintrinsics.Today, however, such loads are normalized to standard
GT_INDon import and the tests validating these intrinsics no longer provide significant additional value, particularly when viewed across the total coverage we have in stress runs, in other real world usage in the libraries, etcAs such, this removes those scenarios from the core test templates so we can decrease the total run time for all hwintrinsic tests.