[PATCH] D38668: [X86] Fix for 20871

[Riyaz V Puthiyapurayil via Phabricator via llvm-commits]

RVP added a comment.

The test you suggested is already there for X86_64 -as mul128.ll.

I can update that test to run for i386 as well. The only problem is that on x86, it generates a huge number of instructions. What should I check for? CHECK-NOT __multi3 or check for the complete instruction sequence? What do you think?

On a different note, I now think that I should not probably position this change as a fix for PR20871. Note that the original test  in that bug report did not have any 128-bit multiplication. I narrowed it down to a combination of licm, instcombine and indvars that ends up generating a 65-bit multiplication which then gets promoted to 128 bits in the selector. I suspect there is a problem upstream that should be fixed independently.


https://reviews.llvm.org/D38668