Toward distributions of native libraries

[Samuel Audet]

On 1/14/19 3:04 PM, John Rose wrote:
> On Jan 12, 2019, at 7:42 PM, Samuel Audet <samuel.audet at gmail.com> wrote:
>>
>> Thanks for the feedback! If it were a priority, and if it were me, I would start by ensuring that the high-level API can also be optimized in the case of AOT, but this isn't happening at the moment. I just shortened it this way because I consider AOT to be just as important as JIT. How do you feel I should phrase this then? That it isn't a priority /only in the case of AOT/? I'll edit it that way if that's accurate! If not, I would welcome additional explanations, for example, see https://github.com/oracle/graal/issues/885 . Thanks
> 
> Performance is important, period.  The discussion on #885 peters out with you saying, "but can it work efficiently?".  That's a complicated question, to which I believe the answer is almost certainly "yes".  We are, fact, planning to make such things "work efficiently", and "such things" include JIT, AOT, and SVM compilation of Panama apps.
> 
> So I agree with Maurizio's comment, that this bit of your blog post (a very good one!) would better reflect our intentions by omitting the comment about performance.  If you wish, you could satisfy our observations by saying something straightforward like "Although the Panama people claim to value high performance, I don't think they can achieve it, especially for AOT."  If that's true for you, in which case, wait and see how it all turns out.
> 
> To put it in a nutshell:  You misrepresent our priorities (a set of mental states Maurizio and I can and do authoritatively explain) by reasoning about their feasibility (which reasonable people can disagree about).
> 

I would rather learn to disagree than to ignore each other, so thank you 
for the feedback!

IMO, the question I asked to the SVM team isn't a complicated one. I 
asked in the simple case of getpid() whether or not we will be able to 
achieve the same speed as "org.graalvm.nativeimage.c", and although you 
may believe the answer is "yes", the fact is that it looks like we don't 
know. So the answer is "probably, but we don't know for sure". What if 
it turns out to be "no"? Are you prepared to refactor the whole 
high-level API to something else entirely and maybe even bring back 
static methods as an option if we're really up against the wall in one 
or more corner cases where they would help performance? If that's the 
way you guys work, that's fine too, but let's be clear about it!

Anyway, I like to stay objective, so I've rephrased the passage on 
http://bytedeco.org/news/2019/01/11/importance-of-a-distribution/ 
without any implications about what you consider important or not:

"numbers showing faster-than-JNI performance for both JIT and AOT 
compilers are still lacking"

Sounds good? I'm happy to hear at least that AOT is of importance too!

Samuel