Re: [PATCH] IPA: support -flto + -flive-patching=inline-clone

["Martin Liška" <mliska@suse.cz>]

On 10/7/22 15:04, Qing Zhao wrote:
> 
> 
>> On Oct 7, 2022, at 2:34 AM, Richard Biener <richard.guenther@gmail.com> wrote:
>>
>> On Thu, Oct 6, 2022 at 3:18 PM Qing Zhao <qing.zhao@oracle.com> wrote:
>>>
>>>
>>>
>>>> On Oct 6, 2022, at 4:29 AM, Richard Biener <richard.guenther@gmail.com> wrote:
>>>>
>>>> On Wed, Oct 5, 2022 at 8:18 PM Qing Zhao via Gcc-patches
>>>> <gcc-patches@gcc.gnu.org> wrote:
>>>>>
>>>>>
>>>>>
>>>>>> On Oct 5, 2022, at 1:36 PM, Martin Liška <mliska@suse.cz> wrote:
>>>>>>
>>>>>> On 10/5/22 16:50, Qing Zhao wrote:
>>>>>>> I have two questions on this:
>>>>>>
>>>>>> Hello.
>>>>>>
>>>>>>>
>>>>>>> 1.  What’s the motivation to enable -flive-patching with -flto? Is there any application that will try -flive-patching with -flto now?
>>>>>>
>>>>>> We're planning supporting GCC LTO Linux kernel support, so that's one motivation. And the second one is a possible
>>>>>> use in user-space livepatching. Note majority of modern distros default to -flto (openSUSE, Fedora, Debian, Ubuntu, ...).
>>>>>
>>>>> Okay, I see. That’s reasonable.
>>>>>>
>>>>>>>
>>>>>>> 2. Why only enable -flive-patching=inline-clone with -flto?
>>>>>>
>>>>>> Because the inline-only-static level (which you added/requested) would have to properly
>>>>>> block inter-procedural inlining that happens in LTO (can_inline_edge_by_limits_p) and
>>>>>> I'm not sure it would be properly blocked. So, feel free to extend my patch if you want?
>>>>>
>>>>> -flive-patching=inline-only-static
>>>>>
>>>>> Only enable static functions inlining,  all the inlining of external visible functions are blocked, So, LTO should be compatible with this naturally without any issue, I think.
>>>>>
>>>>> i.e, when "-flive-patching=inline-only-static -flto"  present together, all the inter-procedural inlining should be automatically blocked by -flive-patching=inline-only-static already.
>>>>>
>>>>> Do I miss anything here?
>>>>
>>>> WPA will promote externally visible functions static when all accesses
>>>> are from LTO IR, I don't think we preserve
>>>> the "original" visibility for IPA inlining heuristics.
>>>
>>> WPA is Whole Program Analysis?
>>
>> Yes.
>>
>>> Okay, then  It will promote all static function to extern functions. That’s reasonable.
>>
>> No, all extern functions to static functions.
> 
> Oh, really? Why change “extern” to “static”?  (I recall that studio compiler promote static to extern for inter-procedural inlining)

Because the linker LTO plug-in can get information about symbols are really
and the thus some of them can before static and not visible.

>>
>>> Is it hard to preserve the original “static” visibility in the IR?
>>
>> Probably not hard, and the IPA pass adjusting visbility could as well
>> mark the functions
>> as not to be inlined with -flive-patching=inline-only-static.
> Okay, then the implementation should be doable?
>>
>>>>
>>>> OTOH inline-only-static could disable WPA inlining and do all inlining early ...
>>>
>>> Inline-only-static ONLY inlines static functions, how can it disable WPA inlining? Don’t quite understand here.
>>
>> it's a flag so it can be used to control other things.
> 
> When -flive-patching=inline-only-static is specified by the user, user explicitly request ONLY inlining static functions.
> Even when LTO is enabled, if only static function inlining is enabled, user gets what he/she want. So I didn’t see any issue here.

Please see Honza's previous email about the second inliner.

Anyway, I'm going to install my patch that handles inline-clone option value.

Martin

> 
> Let me know if I still miss anything here> 
> Thanks.
> 
> Qing
> 
> 
>>
>>> thanks.
>>>
>>> Qing
>>>>
>>>> Richard,
>>>>
>>>>>
>>>>> thanks.
>>>>>
>>>>> Qing
>>>>>
>>>>>>
>>>>>> Martin
>>>>>>
>>>>>>>
>>>>>>> thanks.
>