Re: Question on patch -fprofile-partial-training

[Qing Zhao <qing.zhao@oracle.com>]

> On May 4, 2023, at 9:05 AM, Martin Liška <mliska@suse.cz> wrote:
> 
> On 5/4/23 14:54, Qing Zhao wrote:
>> 
>> 
>>> On May 4, 2023, at 4:30 AM, Martin Liška <mliska@suse.cz> wrote:
>>> 
>>> On 5/3/23 21:10, Qing Zhao via Gcc-patches wrote:
>>>> Hi, Jan,
>>>> 
>>>> You added the following patch into gcc10:
>>>> 
>>>> From 34fbe3f0946f88828765184ed6581bda62cdf49f Mon Sep 17 00:00:00 2001
>>>> From: Jan Hubicka <hubicka@ucw.cz>
>>>> Date: Thu, 5 Dec 2019 19:12:51 +0100
>>>> Subject: [PATCH] cgraphclones.c (localize_profile): New function.
>>>> 
>>>>       * cgraphclones.c (localize_profile): New function.
>>>>       (cgraph_node::create_clone): Use it for partial profiles.
>>>>       * common.opt (fprofile-partial-training): New flag.
>>>>       * doc/invoke.texi (-fprofile-partial-training): Document.
>>>>       * ipa-cp.c (update_profiling_info): For partial profiles do not
>>>>       set function profile to zero.
>>>>       * profile.c (compute_branch_probabilities): With partial profile
>>>>       watch if edge count is zero and turn all probabilities to guessed.
>>>>       (compute_branch_probabilities): For partial profiles do not apply
>>>>       profile when entry count is zero.
>>>>       * tree-profile.c (tree_profiling): Only do value_profile_transformations
>>>>       when profile is read.
>>>> 
>>>> My question is:
>>> 
>>> Hello.
>>> 
>>> Why would anybody backport such change to unsupported code-stream of GCC 8?
>>> Generally speaking, I discourage from doing that.
>> 
>> Yes, I agree.
>> However, many users still use GCC8 right now, and some of them are asking for more performance
>> from PGO recently. That’s the reason I am studying this right now. 
> 
> I understand there are products that are based on GCC8, but as the branch is officially unsupported, I don't
> see a reason to backport a new feature from newer release. It's just asking for troubles. If your clients are
> interested in more performance, then they should use a recent supported release.
We are trying to persuade them to use newer GCC, but it’s quite hard...
> 
>> 
>> From my understanding, -fprofile-partial-training is one important option for PGO performance.
> 
> I don't think so, speed benefit would be rather small I guess.
I saw some articles online to introduce this option for gcc10, 
https://documentation.suse.com/sbp/all/html/SBP-GCC-10/index.html#sec-gcc10-pgo
And also based on my previous experience in Studio compiler, I guess that this one might have
Some good performance impact on PGO.  Is there any old performance data on this option? (I cannot find online)

thanks.

Qing

> 
>> I’d like
>> to see any big technique difficult to prevent it from being back ported to GCC8. 
> 
> There might be of course some patch dependencies and I don't see a point why should we waste
> time with that.
> 
> Cheers,
> Martin
> 
>> 
>> Thanks.
>> 
>> Qing
>> 
>>> 
>>> Martin
>>> 
>>>> 
>>>> Can this patch be back ported to GCC8 easily? I am wondering any significant
>>>> Change between GCC8 and GCC10 that might make the backporting very hard> 
>>>> Thanks a lot for your help.
>>>> 
>>>> Qing