[PATCH] D19087: [x86, ppc] prefer comparisons against zero for and+cmp sequences

[Sanjay Patel via llvm-commits]

spatel added a comment.

In http://reviews.llvm.org/D19087#401155, @hfinkel wrote:

> In http://reviews.llvm.org/D19087#401119, @spatel wrote:
>
> > In http://reviews.llvm.org/D19087#400599, @hfinkel wrote:
> >
> > > This certainly a good pattern to catch, but I'm not sure that CGP is the right place for this. We generally have things in CGP to work-around the fact that SDAG/ISel is basic-block local. This kind of thing seems much more natural as a something that should be in DAGCombine.
> >
> >
> > Yes, I initially thought I'd do this as a DAGCombine, but then I realized that we'd need "isKnownToBeAPowerOfTwo()", and I don't see an SDNode equivalent for the IR version. I'm not sure where we draw the line between CGP and DAGCombine, but duplicating isKnownToBeAPowerOfTwo() scared me away.
>
>
> Why? How do we select the instructions anyway? For x86 or ppc, would we not want to limit the applicability to constants? It needs to be a constant for rlwinm I'd imagine, and bt too?. For constants, we can check this with existing functionality.


isKnownToBeAPowerOfTwo() is ~100 lines and recursive. We can approximate it using computeKnownBits(), but I think getting the full power of the IR version would mean duplicating the code.

One of the x86 regression tests shows the case where a simple constant check won't do:

  define i1 @and_cmp_const_power_of_two(i32 %x, i32 %y) {
    %shl = shl i32 1, %y
    %and = and i32 %x, %shl
    %cmp = icmp ne i32 %and, %shl
    ret i1 %cmp
  }

X86TargetLowering::LowerToBT() uses computeKnownBits() to know that's a power-of-2 mask.


http://reviews.llvm.org/D19087