[LLVMdev] A question about everyone's favorite constructs: NSW and NUW

[Quentin Colombet]

Hi Chandler,

On Jan 21, 2014, at 9:53 AM, Chandler Carruth <chandlerc at google.com> wrote:

> I've not yet looked at the code, but just to give you somewhat raw numbers:
Thanks a lot for your measurements!

> 
> Across all of our benchmarks there was overall change outside of the noise. There is one benchmark that may have been improved, but I'm not 100% certain it'll be reproducible as opposed to a fluke run.
That was expected as this patch is intended to help code that runs on 64-bit architectures that is not 64-bit friendly for array accesses. I believe most of your code base is not in that shape :).

> 
> While this isn't fantastic news, on the flip side, nothing regressed, so that's a really good sign. Also, no size regressions to speak of (0.02%? yea, no one will care).
That’s good news because on our side we are seeing good performance improvement in the llvm test-suite and some external benchmarks like SPEC.
> 
> From a performance stance I think this patch looks fine if you guys are seeing significant improvements.
I am still investigating our numbers because we are seeing some regressions that I need to confirm (so far, I saw only false positive).

Anyway, thanks again!
-Quentin 
> 
> 
> On Mon, Jan 20, 2014 at 11:58 AM, Quentin Colombet <qcolombet at apple.com> wrote:
> Hi Chandler,
> 
> Here is an updated patch, that I plan to submit for review.
> In particular, it has the rollback mechanism when the addressing mode is not profitable.
> 
> Could you run some benchmarks with it?
> 
> Thanks,
> -Quentin
> 
> 
> On Jan 7, 2014, at 11:40 AM, Chandler Carruth <chandlerc at google.com> wrote:
> 
>> 
>> On Tue, Jan 7, 2014 at 2:24 PM, Quentin Colombet <qcolombet at apple.com> wrote:
>> I agree with Andy and Jim.
>> Actually, I have prototyped a compiler that does exactly this kind of promotion in CodeGenPrepare.
>> Basically, I have updated the addressing mode matcher so that it moves a sext that is in a way of an addressing mode (i.e., it promotes the operand of the sext, let us call this operand def, if it is legal to do so, and sign extends the operands of def). When the matcher does not manage to absorb more computation after promoting def, it can revert the promotion.
>> 
>> I am currently benchmarking this solution and I’ll update this thread with the results.
>> 
>> Very cool. Could you share the patch? I can also run some benchmarks.
> 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140121/508cbe81/attachment.html>