[llvm-commits] x86 branch sequence optimization in LLVM code gen: please review

[Evan Cheng]

On Jan 5, 2012, at 2:01 AM, Chandler Carruth wrote:

> On Thu, Jan 5, 2012 at 1:53 AM, Umansky, Victor <victor.umansky at intel.com> wrote:
> Well, if it would be possible to change LLVM intrinsic prototype – then you’re right, no instruction combining were required.
> 
> However until this will be done in LLVM spec – the optimization is in place.
> 
> 
> There is no spec, it is a change to LLVM's intrinsics, you can post a patch to do this to this mailing list. =] You need to change the intrinsic definition, update any code that reasons about that intrinsic, add auto-upgrade logic to upgrade old IR inputs from the old intrinsic to the new intrinsic, and teach the frontends to emit the new one. It's a fair amount of code, but none of it hard.
> 
> Please don't leave bad code checked into the codegen layer. We need to fix this the correct way if/when we find the correct way to fix it. Thus far, updating the LLVM intrinsics is looking very promising.

I agree with Chandler. Victor, can you follow his suggestion and fix this properly? Sorry, I didn't think it through when I was reviewing the patch.

Evan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120105/f35cb79f/attachment.html>