Re: [Bug-apl] Optimizations revived

[Juergen Sauermann]

Hi Elias, there are actually more cases. For example:       (5 5↑A)←B This is like left values in C/C++. I would agree that unnecessary cloning should be avoided. But that needs to be done in a way that can be proven to be correct (and without complex data structures managing ownership of values). In the past even a rather simple flag scheme has turned out to be non-maintainable and I would not like to walk on such a path again. You can have a look at tags/apl-1.1-old-memory-management/ in SVN to see how to not do things. /// Jürgen On 08/24/2014 05:01 PM, Elias Mårtenson wrote: Of course, there are definitely times where cloning is necessary. The typical example would of course be when a value is assigned to another and that value is subsequently changed (by using indexed assignment). And you mentioning the issue of not know who are the owners of an object reminds me of the fact that I also came across that issue. At first I was thinking that the problem is trivial; that you simply clone if the owner count is >1. Then I realised that there could be many owners, but those owners are not going to touch the object anymore. There must be an optimal solution here. I'm sure you agree with me about the ideal situation: Where there is no cloning except for when it's needed because an object will be changed. Now: Can you confirm if I'm correct in my assessment that the only thing that forces cloning in the first place is indexed assignment? If APL had no such feature, then there would be zero need for cloning, yes? Regards, Elias On 24 August 2014 22:46, Juergen Sauermann <juergen.sauerm...@t-online.de> wrote: Hi Elias, I don't know. Actually I just found a nasty example which renders my statement below ("I believe the clone() call in Symbol::resolve() can be skipped completely") wrong:       A←10⍴10       A+ (A[3]←3) ⊢ A Currently a value is cloned when it is assigned to a variable and again when the variable is referenced. The above example shows why cloning on reference is needed. You get the correct 20 20 13 20 20 20 20 20 20 20 when you clone and the wrong 20 20 6 20 20 20 20 20 20 20 if not. The problem that I see with CoW is that we know how many owners a value has but not who they are. And you definitely dont want to clone on indexed assignment because very often you have a large value and only a small part being updated. With CoW you dont want to clone the value that is being assigned but rather the other (possibly many copies) and those we don't know. I also don't share your opinion that the cloning is unnecessary. I think it is more so that it is sometimes unnecessary, but at the time when you would not  clone you don't have enough information to properly make that decision. /// Jürgen On 08/24/2014 04:04 PM, Elias Mårtenson wrote: Thanks for the explanation. With this in mind, I'm wondering about two specific cases and how you intend to solve them (these are are ones that gave me the biggest headache): First of all, how will you avoid doing a clone() on every update using indexed assignment? You need some way to know whether or not an instance have to be cloned or can be modified in-place. A related case is normal left-arrow assignment. It is not desirable to do a clone() on every assignment either. You want to basically do a "deferred clone" (i.e. CoW) so that the instances can be shared in both variables until one of them is updated. When I was thinking about this, I couldn't get away from concluding that a full CoW implementation is needed in order to avoid all unnecessary cloning. Regards, Elias On 24 August 2014 00:01, Juergen Sauermann <juergen.sauerm...@t-online.de> wrote: Hi Elias, normally APL values are not written to. The exception is indexed assignment. I believe the clone() call in Symbol::resolve() can be skipped completely. This is probably the most frequently used clone() case. I suppose copy-on-write semantics is achieved when all clone() calls are gone. Many of the remaining clone() calls are specific to certain functions so their performance impact should be small. I haven't done the above before the 1.4 release because I didn't want to release a not-so-well tested optimization. /// Jürgen On 08/23/2014 05:32 PM, Elias Mårtenson wrote: Cool :-) Speaking of this, I did spend quite a bit of time some time ago to try to figure out an easy way to get generic copy-on-write semantics, but I never came to a satisfactory conclusion. I might revisit it later. Have you thought about it? It is, after all, related to this specific optimisation. Regards, Elias On 23 August 2014 23:29, Juergen Sauermann <juergen.sauerm...@t-online.de> wrote: Hi, I have revived Elias' in-place optimization for A⍴B and ,B now usinng a different way of figuring if B is in use. SVN 445. /// Jürgen