[PATCH] D122077: [InstCombine] Fold (ctpop(X) == 1) | (X == 0) into ctpop(X) < 2

Tue Mar 29 05:13:36 PDT 2022

spatel added a comment.

In D122077#3413565 <https://reviews.llvm.org/D122077#3413565>, @xbolva00 wrote:

> In D122077#3413550 <https://reviews.llvm.org/D122077#3413550>, @joerg wrote:
>
>> Why is this fold preferable to `(X & (X-1)) == 0`? At least on all architectures without native population count, the binary-and based test is preferable and it might even be better with it.
>
> Less IR instructions. I think SDAG already expands some ctpop patterns to logic (backends should decide about optimal form)

Correct - we try to convert the setcc(ctpop) pattern here:
https://github.com/llvm/llvm-project/blob/f1d8e46258c6a08ca1a375dc9670dd5581d6cf65/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp#L3772

That doesn't check whether the target has a popcount instruction for scalar types, so it is potentially too aggressive. For x86, we have a bonus optimization after that, so it should not show up there:
https://godbolt.org/z/oc16Kdhcf

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D122077/new/

https://reviews.llvm.org/D122077