Alexander Root

@steven-johnson Any chance that #7083 works now?

Yes, it should be fine - the phadd intrinsics have been in LLVM for over a decade [sse3](https://github.com/llvm/llvm-project/commit/2377206923c1b77296899b44a77ae814806ff2f2) [avx2](https://github.com/llvm/llvm-project/commit/682b850602d3380528c5ecafe352c48422e19724).

Failures were due to a shuffling mistake, I forgot that AVX2 has the 128 bit boundaries for instructions - fixed in 0acae70, and confirmed that we still produce better codegen...

We don't have any existing patterns for `phadd` instructions, so I'll only respond to your question in regards to the use of `pmaddwd` here. I did not consider rewriting within...

I will say that I don't think these particular patterns are suitable for rewriting to enable `should_use_dot_product`, but instead I will aim for producing the patterns that will match the...

I might close this PR in favor of #6884 , which will make the `horizontal_widening_add` patterns useful on any architecture that has `pmaddwd` variants. Just waiting to get feedback on...

Yes, I am closing this in favor of #6884 . If we end up not merging #6884 for some reason, I will re-open this PR.

@ryanstout if you give incorrect estimates (and the actual size is smaller than the estimates), I believe it's possible for the tiling generated by the autoscheduler to access out of...

@dsharletg Did you fix this somehow? This reproducer no longer fails on current master for me, nor does adams2019 on bgu (which has been consistently failing due to the same...

I checked my recent changes, and the commit before them still doesn't trigger this reproducer, I'm going to check if the is_monotonic bug was the issue