[LLVMdev] RFC - Improvements to PGO profile support

[Xinliang David Li]

On Tue, Mar 24, 2015 at 1:13 PM, Chandler Carruth <chandlerc at google.com>
wrote:

>
> On Tue, Mar 24, 2015 at 1:08 PM, Xinliang David Li <xinliangli at gmail.com>
> wrote:
>
>> Example. Assuming the cap is 'C'
>>
>> void bar()
>> {
>>     // ENTRY count is 4*C, after capping it becomes 'C'
>>     ...
>> }
>>
>> void test()
>> {
>>   // BB1:   count(BB1) = C
>>   bar();
>>
>>   // BB2:   count(BB2) = C
>>   bar();
>>
>> }
>>
>> void test2()
>> {
>>   // BB3:   count(BB3) = C
>>   bar();
>>
>>   // BB4:   count(BB4) = C
>>   bar();
>> }
>>
>> What would inliner see here ? When it sees callsite1 -- it might mistaken
>> that is the only dominating callsite to 'bar'.
>>
>
> I don't understand, why would it assume this?
>

because

count(ENTRY_bar) = SUM (count(callsite_i(bar)), for all i.

If the inliner sees the first callsite has the same count as the bar's
entry count, what assumption can it make (especially when the module has
only one callsite to be looked at)? With more callsites visible, it can
detect the insanity, but still it makes the data less useful.


>
> I was suggesting that each function would need to be associated with some
> global weight, and the global weight ration between test and bar should
> provide the needed information here?
>

See my previous reply. The data is there. We just need to preserve it.

David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150324/fc62e2f0/attachment.html>