Re: Sparse 0.5.1 RC5 released.

[Dibyendu Majumdar <mobile@majumdar.org.uk>]

Hi Luc,

On 13 August 2017 at 15:25, Luc Van Oostenryck
<luc.vanoostenryck@gmail.com> wrote:
>> OTOH with -rc4, it takes only 0.05s !
>>
>> The situation is really weird, it's an interaction between the patches:
>> - remove single-store shortcut (first slowdown to 0.8s)
>> - fix ptrlist corruption while killing unreachable BBs (second slowdown to 2.9s)
>> - mark pseudo users as deleted instead of removing them (third slowdown to 5.5s)
>>
>> All three patches were expected to create *some* slowdown,
>> the first one very obviously and much more than the others.
>>
>> It's interesting to note that without the 'remove single shortcut' patch, the
>> two others ones don't seem to have a significative effect on the time.
>>
>> My guess here is that indeed, with this input, the single-store shortcut
>> is a huge optimization and without it much much more code is created
>> (phi-nodes, I suppose) and this impact very negatively the two other fixes.
>>
>> I'll see what can be done.
>
> It's also interesting to note that with the new SSA construction, there is no
> slowdown (0.064s here but too small to be measured accurately).
> This confirms me that the single-store shortcut was very much needed for
> performance (not simply because it's faster than the general case but that
> it creates much less intermediate code) but it created bad code ...
>

I wonder if it is worth doing these changes - instead we should
probably look at your new SSA construction approach.

Regards
Dibyendu