On-Demand range technology [4/5] - Performance results

[Andrew MacLeod]

On 5/23/19 9:29 AM, Richard Biener wrote:
> On Thu, May 23, 2019 at 3:29 AM Andrew MacLeod <amacleod@redhat.com> wrote:
>> All times are with a release configured compiler.  Out of the 242 files,
>> pretty much across the board in all 4 sets of figures, RVRP was faster
>> in about 90% of the cases, and slower in the other 10%, resulting in the
>> following cumulative totals.
>>
>>                               Overall (242 files)
>> 1 - Raw RVRP                              22% slower   (**a**)
>> 2 - No edge calculation(1)                4.5% slower
>> 3 - No switch processing(2)                9.5% faster
>> 4 - No dominators(3)                    16% faster  (**b**)
>> 5 - Conditionals (including switches) only        4% faster
> I don't understand a vs. b here.  Does this mean dominators
> cost 38% performance?!
>
> Does 4 imply 3 and 2 maybe?
>
> Dominators cannot be this expensive.
>
> hope

I believe 4 also has no switch processing, so yes, it implies 3 which 
also includes 2.   so we go from 10% faster to 16% faster without 
building the dominators.  Probably more the kind of number you were 
expecting :-)

Andrew