[llvm] [RISCV] Move strength reduction of mul X, 3/5/9*2^N to combine (PR #89966)

[Philip Reames via llvm-commits]

preames wrote:

@dtcxzyw Thank you for finding and reporting this code gen difference.  And also, yuck!

This is actually the same issue I hit in X86 when trying to use the generic SHL_ADD node, so I've already investigated it a bit.  In short, it really doesn't look like we have any good answers here.  There's existing discussion around this here: https://github.com/llvm/llvm-project/pull/84921#discussion_r1523719273

My takeaway is that freeze in DAG is only half implemented, and there's really no clear good path forward.

I very reluctantly think it may be time to give up here.  No other target is actually undef correct for these cases, so maybe we shouldn't bother to be either.  I've posted a patch for that here: https://github.com/llvm/llvm-project/pull/90097   If by chance anyone has a better idea, I'm definitely open to it.  

https://github.com/llvm/llvm-project/pull/89966