[PATCH] D54115: [InstCombine] do not shrink switch conditions to illegal types (PR29009)

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

To provide more background/motivation for why I don't think we will solve the underlying problem soon, we can fork this example into its own bug:
https://bugs.llvm.org/show_bug.cgi?id=29009#c4

I said that was an instcombine opportunity in that bug comment, but that was a mistake. It probably requires a GVN-level-of-complexity to see that kind of equivalence between the 2 different-sized masked values. But AFAICT, neither -gvn nor -newgvn optimizes that today (so we should file a 2nd bug to track that as an IR opportunity).

I'm not sure where/how we would catch that in the backend, but that's the kind of code we would see after legalization of non-power-of-2 types.

So that's the reason I think we're justified in limiting this transform with shouldChangeType(). In an ideal world, we could let instcombine go wild with illegal types, but as this example shows, getting optimal codegen from that is likely difficult (and it might not be handled well by other IR passes too), so that's why instcombine has this limitation.


Repository:
  rL LLVM

https://reviews.llvm.org/D54115