[PATCH] D32143: [InstSimplify] use ConstantRange to simplify more and-of-icmps

[Ulrich Weigand via Phabricator via llvm-commits]

uweigand added a comment.

I'm seeing a performance regression which I traced back to this commit.   A reduced test case to show the problem is this:

  define i1 @test(i64 %a) {
    %cmpge = icmp sge i64 %a, 1
    %cmpeq = icmp eq i64 %a, 1
    %cmpne = icmp eq i1 %cmpeq, false
    %cmp = and i1 %cmpge, %cmpne
    ret i1 %cmp
  }

Before this commit, InstCombine would reduce the above to simply:

  define i1 @test(i64 %a) {
    %1 = icmp sgt i64 %a, 1
    ret i1 %1
  }

but for some reason this no longer happens now.

Note that when directly using an "icmp ne" instead of inverting the result of the "icmp eq", the optimization does still happen.


Repository:
  rL LLVM

https://reviews.llvm.org/D32143