[llvm-dev] Does anyone know a workaround for this branching bug I hit?
Craig Topper via llvm-dev
llvm-dev at lists.llvm.org
Mon Apr 27 00:27:44 PDT 2020
You can try "-mllvm -x86-cmov-converter=false". That should keep it as a
cmov, but might give you performance issues elsewhere.
On Mon, Apr 27, 2020 at 12:14 AM Levo DeLellis via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> I believe you all can access the test file and the diff in my bug report
> In short, I'm experimenting with optimization and wrote branchless code
> -O2 generates a jump instruction causing my loop to be twice as slow. I
> changed the jne+movl to cmove and my loop seems to take half as long.
> I'm a little worried that this will interfere with other optimizations I
> make in the same function (bug report is a small reproduce). Is there a
> temporary workaround I can use? I'm thinking no and that I would have to
> suck it up. But I seen stranger things so I thought I'd ask. I guess for
> now I'll experiment allowing the slow jumps and I'll hand change it once
> I'm done optimizing that function
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev