Re: moveFront() and friends: Request for comment
On 2010-08-27 14:04:50 -0400, Andrei Alexandrescu said: Add to this the implementation annoyance of checking for aliasing in _every_ single method of the type. RefCounted in phobos can help with that, but not without a cost. Also, to get reference counting right when the struct can be stored on the garbage-collected heap you must use atomic operations to manipulate the reference count. Atomic operations add some more overhead on multi-core and multi-processor systems. -- Michel Fortin michel.for...@michelf.com http://michelf.com/
Re: moveFront() and friends: Request for comment
== Quote from Andrei Alexandrescu (seewebsiteforem...@erdani.org)'s article > On 8/27/10 7:44 PDT, Pillsy wrote: > > dsimcha Wrote: > > [...] > >> I'd like any comments anyone might have on to what extent > >> arbitrarily expensive postblits should be considered in the > >> design of Phobos. > > > > I agree with you: expensive post-blits just don't seem sufficiently > > necessary in D to warp the design of the standard library around > > them, and have a distinctly anti-patternish feel to them. You have > > too many other options, like straight reference semantics, copy-on- > > write, and (with immutability and GC) safely shared structure. > > > > The last can be an incredibly useful technique for reducing the > > cost of copying, because it allows you to treat an immutable > > reference type exactly like a value type. > > > > Cheers, > > Pillsy > Clearly immutable sharing helps, and clearly reference counting and COW > are valid techniques. However, the situation is not as cut and dried. > The problem with RC/COW is that they reduce exception safety essentially > _everywhere_ else but the copy constructor. Once you have a type using > RC/COW, any mutation of that object, even one that ostensibly doesn't > cause resource allocation, might have arbitrary cost or fail. This was > the experience with std::string in C++ - its creators did everything > they could to enable reference counting, and the outcome was quite > unpleasant. > Add to this the implementation annoyance of checking for aliasing in > _every_ single method of the type. RefCounted in phobos can help with > that, but not without a cost. > So I'm not sure it's as simple a decision as it might sound. > Andrei But with RC/COW at least you know that if something becomes arbitrarily expensive it's for a good reason, i.e. because doing anything else would break the abstraction you're trying to create **right now**. With eager copying copies get made very often when a human reader reasoning only locally about the code could easily prove they're not needed. Furthermore, I think my point still generally holds about reference vs. value semantics: When you have to bend over backwards to get value semantics, they're no longer easier to reason about.
Re: moveFront() and friends: Request for comment
On 8/27/10 7:44 PDT, Pillsy wrote: dsimcha Wrote: [...] I'd like any comments anyone might have on to what extent arbitrarily expensive postblits should be considered in the design of Phobos. I agree with you: expensive post-blits just don't seem sufficiently necessary in D to warp the design of the standard library around them, and have a distinctly anti-patternish feel to them. You have too many other options, like straight reference semantics, copy-on- write, and (with immutability and GC) safely shared structure. The last can be an incredibly useful technique for reducing the cost of copying, because it allows you to treat an immutable reference type exactly like a value type. Cheers, Pillsy Clearly immutable sharing helps, and clearly reference counting and COW are valid techniques. However, the situation is not as cut and dried. The problem with RC/COW is that they reduce exception safety essentially _everywhere_ else but the copy constructor. Once you have a type using RC/COW, any mutation of that object, even one that ostensibly doesn't cause resource allocation, might have arbitrary cost or fail. This was the experience with std::string in C++ - its creators did everything they could to enable reference counting, and the outcome was quite unpleasant. Add to this the implementation annoyance of checking for aliasing in _every_ single method of the type. RefCounted in phobos can help with that, but not without a cost. So I'm not sure it's as simple a decision as it might sound. Andrei
Re: moveFront() and friends: Request for comment
dsimcha Wrote: [...] > I'd like any comments anyone might have on to what extent > arbitrarily expensive postblits should be considered in the > design of Phobos. I agree with you: expensive post-blits just don't seem sufficiently necessary in D to warp the design of the standard library around them, and have a distinctly anti-patternish feel to them. You have too many other options, like straight reference semantics, copy-on- write, and (with immutability and GC) safely shared structure. The last can be an incredibly useful technique for reducing the cost of copying, because it allows you to treat an immutable reference type exactly like a value type. Cheers, Pillsy