[llvm-dev] DAGCombiner folds (xor (and x, y), y) -> (and (not x), y). Can I do the reverse for a target?

[Assem Bsoul via llvm-dev]

In SystemZ, for 64-bit operands, (not x) is translated as two xor
instructions, i.e., XORing the high and low half words of x with 32-bit
immediate. The other form that I want will reduce the instruction count by
1 because (and x, y) is translated as a single instruction.

PS. I am copying the below because I am not sure if it will be included in
the reply by default

Thanks
Assem

On Tue, Jun 21, 2016 at 12:49:14PM -0400, Assem Bsoul via llvm-dev wrote:
> DAGCombiner currently folds  (xor (and x, y), y) -> (and (not x), y)
>
> I was trying to do the reverse of this transformation, i.e., (and (xor x,
> -1), y) -> (xor (and x, y), y), and recognized that this causes an
infinite
> loop. What's the advantage of doing the above folding? if I want to do
the
> reverse for a specific target where I can place that and make sure it
> doesn't get reversed by the above folding?

The goal of the transform is to provide a simpler canonical form. E.g.
(not x) can be computed independently and it is also easier to reason
about. Why is it beneficial for your target to have the first form?

Joerg
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160621/7645af68/attachment.html>