[PATCH] D28966: [PGO] Memory intrinsic calls optimization based on profiled size

[Rong Xu via llvm-commits]

OK. I will split this patch into two: One for annotation and the other for
the transformation.

On Wed, Mar 8, 2017 at 11:42 AM, Xinliang David Li <davidxl at google.com>
wrote:

>
>
> On Wed, Mar 8, 2017 at 11:29 AM, Rong Xu <xur at google.com> wrote:
>
>>
>>
>> On Wed, Mar 8, 2017 at 11:24 AM, Xinliang David Li <davidxl at google.com>
>> wrote:
>>
>>>
>>>
>>> On Wed, Mar 8, 2017 at 11:11 AM, Rong Xu via Phabricator <
>>> reviews at reviews.llvm.org> wrote:
>>>
>>>> xur added a comment.
>>>>
>>>> In https://reviews.llvm.org/D28966#694902, @davidxl wrote:
>>>>
>>>> > The optimization pass should be split into two phases as IC
>>>> promotion. The annotation part should probably  be merged with the
>>>> instrumentation patch. The transformation patch should be done in the same
>>>> pass as IndirectCallPromotion.
>>>>
>>>>
>>>> Is there a good reason for doing the transformation late?  Here I don't
>>>> do the annotation and instead, I do the transformation directly in the same
>>>> pass. The main reason we have annotations in indirect-call-promotion is we
>>>> need to call it late (in LTO or ThinLTO). Another reason I'm reluctant to
>>>> use annotation is that we need to maintain/update it (for inline and
>>>> unroll, for example).
>>>>
>>>
>>> Unlike the indirect call promotion, which enables more inlining,
>>> stringOp value prof transformation can lead to increased function body size
>>> that negatively affect inlining decisions.
>>>
>>> This is true. But the profile we collect is before inline.  The count
>> values does not have context information. The only thing we can do is to
>> scale down based on the edge count.
>>
>
> Right, but there is no profile meta-data update needed either.  The value
> counts total should match the enclosing BB count. If not, scale it before
> deciding profit.  Even though it does not have full context sensitivity, it
> provides some benefit of context hotness. When the call context const prop
> the size parameter, the transformation can be skipped as well.
>
>
>>
>> This also required a separated pass -- we don't have this for
>> indirect-call promotion either (the LTO part of the ICP is different).
>>
>
> Indirect Call Promotion is a separate pass.
>
> David
>
>
>
>>
>>
>>> David
>>>
>>>
>>>>
>>>> -Rong
>>>>
>>>>
>>>> https://reviews.llvm.org/D28966
>>>>
>>>>
>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170308/8e0ab2e0/attachment.html>