As an old C/C++ programmer, I'm always surprised by how often software developers are surprised by the performance costs of inopportune value semantics (C and C++ even more so, punishes you severely for using value semantics when you shouldn't). I increasingly see the wisdom of languages with implicit reference semantics.
It's not that value semantics can't be better (they most assuredly can be), or that reference semantics don't cause their own complexity problems, but rather that so often we thoughtlessly imply/impose value semantics through interfaces in ways that negatively impact performance; getting interfaces wrong is a much tougher bell to unring.
The vast majority of my mental energy when I define an interface in C++ is carefully thinking through a combination of ownership contracts and value vs. reference semantics that I can mostly ignore in languages with implicit reference semantics. While occasionally ignoring those contracts while developing in Java/Python/whatever comes back to bite me, the problem isn't nearly as common or problematic as when I unintentionally impose value semantics in a language that allows me to.
> I increasingly see the wisdom of languages with implicit reference semantics.
I spend most of my time in a JVM language of one flavor or another, and when I was learning Go, the first thing that stuck out at me was, "why would I ever want the compiler to invisibly copy a data structure for me?"
I suppose the primary reason is to prevent the callee from modifying the caller's data out from under them; unless you pass a reference value, you know the callee cannot modify your data.
But, as someone who leans heavily into "everything should be as immutable as possible," the second thing that stuck out at me was "wait, a struct can't have const fields?"
When I write code, it's common to have references to immutable classes thrown around with wild abandon, heedless of ownership, threads, or good taste, because the data just can't change. But that's a paradigm that Go simply doesn't support.
> When I write code, it's common to have references to immutable classes thrown around with wild abandon, heedless of ownership, threads, or good taste, because the data just can't change.
If there's anything I wish languages with implicit reference semantics would adopt, it's implicit immutability. I wish Java would be so much nicer with keyword that is half way between "final" and "volatile" that means, "yes, you can actually mutate this" and then make final semantics the default for fields & variables.
Could you Imagine a Java where you have a `Map` and a `MutableMap` and that's what you put at your API? I'd make it SO much clearer how safe any individual API is to call.
Scala has had this for ages. You can have it today. Even in Java either through the Google collection library or through a library that mimics fp style programming. The name eludes me for the moment.
In general this doesn't work, the history rule says mutable types are not proper subtypes of immutable ones (and the converse is obvious). If you want to capture mutability in your type system, it needs to be orthogonal to subtyping (like C/C++ const).
C++ still has this problem - std::unordered_map<std::string, std::string>` and `std::unordered_map<std::string, const std::string>` are basically unrelated types - you can't const-cast the templated const away. (I may be misunderstanding here)
Yup, it's definitely a bit of a code smell if you do. The issue is more the reverse, though - I can't make a mutable map, then hand it by pointer/reference to something that says it wants an immutable map later.
That only matters if you're storing it. The big issue is 99% of code out there uses mutable template types for containers, and if you ever declare a container that doesn't have a mutable template type, you stub your toe as your new container isn't compatible with anyone else's code. You can't even easily copy your way out.
std::unordered_map<std::string, const std::string> m;
m["foo"] = "bar";
std::unordered_map<std::string, std::string> m2 = m;
Or if you (or something you call) mutate it trough another reference/pointer. Sometimes I wish there was a stronger const where the compiler (and programmer) could actually assume the object is not mutated while that reference/pointer is alive. Of course checking that is no longer doable with just the type system.
> You can't even easily copy your way out.
That depends on the type. std::unordered_maps could provide a copy constructor that does the conversion.
And you can do the copy via iterators, although that is probably not as efficient as it could be because it needs to re-hash the keys.
No, the primary point of a const cast is "I need to pass this to an API that expects a non-const pointer even though it won't mutate it".
Casting const away and then mutating is a footgun as you are in undefined behavior territory as soon as your pointer/reference is not just const itself but points to a const variable.
I had to look it up too, it apparently is a constraint for subtypes defined in Liskovs substitution principle [1]. From Wikipedia:
> History constraint (the "history rule"). Objects are regarded as being modifiable only through their methods (encapsulation). Because subtypes may introduce methods that are not present in the supertype, the introduction of these methods may allow state changes in the subtype that are not permissible in the supertype. The history constraint prohibits this. It was the novel element introduced by Liskov and Wing. A violation of this constraint can be exemplified by defining a mutable point as a subtype of an immutable point. This is a violation of the history constraint, because in the history of the immutable point, the state is always the same after creation, so it cannot include the history of a mutable point in general. Fields added to the subtype may however be safely modified because they are not observable through the supertype methods. Thus, one can define a circle with immutable center and mutable radius as a subtype of an immutable point without violating the history constraint.
Just curious, isn't it obvious from a logical standpoint?
I don't see how one could consider a mutable type to be a subtype of an immutable one. On the other hand, an immutable subtype of a mutable one seem plausible?
Immutable from a mutable is just as implausible. You cannot remove a method from a subtype which results in mutating methods being disabled through other means (like throwing exceptions).
This is also a violation of LSP and the open close principle.
Consider a `List` with an `add` function. What would you do with that `add` function to make an `ImmutableList` subtype?
Ah I see.
Why throw exceptions? They could just be noop depending on the typestate?
But still it doesn't look very nice in practice. I think that in that case, it does make much more sense for a mutable type to be a subtype of the immutable one.
The immutable threw me off track. It's just that the supertype would lack mutation operations. (in a sense immutable would be the default for every type).
Conceptually, you can have constness in your subtype-system (as long as you are sticking with interfaces (methods), as Liskov's subtyping model does, and aren't inherting potentially-mutable fields).
MutableMap and ImmutableMap are both subtypes of a hypothetical ReadableMap. ImmutableMap is the same as ReadableMap, but has an informal contract that subclasses shouldn't add mutability.
Kotlin has this, but the Map is (usually) a MutableMap under the covers, because it's Java bytecode at the lower levels. You have to go out of your way to footgun yourself, but it's still possible.
> When I write code, it's common to have references to immutable classes thrown around with wild abandon, heedless of ownership, threads, or good taste, because the data just can't change. But that's a paradigm that Go simply doesn't support.
You might get a kick out of Virgil. It's easy (and terse!) to define immutable classes and you can have immutable ADTs too. (plus tuples, generics with separate typechecking, etc).
I also have a background in C/C++, etc and I've only ever found myself missing value semantics when I use languages with implicit reference semantics. I guess I always figured the solution was "value semantics with better education / tooling". Education: people should understand value semantics. Tooling: imagine an IDE that highlights allocation points automatically (or perhaps the problem is implicit allocations rather than value semantics?).
> perhaps the problem is implicit allocations rather than value semantics?
I think that’s true. Expensive copies should never have been implicit. There was a story some time ago about a single keypress in the address bar of Chrome causing thousands of memory allocations. The culprit: lots of std::string arguments up and down the call stack.
Rust gets this right, with the hindsight of C++’s example: “a = b” is a move operation by default and clone() is always explicit, except for plain data types where copying is literally memcpy — and those are clearly marked as such by the type system.
IMHO, implicit allocations is a bit of a red herring. Yes, in C/C++ heap allocations are proportionately pretty expensive, but I've seen Java programs have just ridiculous amounts of implicit allocations but there really isn't much of a problem.
But allocations aren't the same as copies, and the argument for reference semantics has always been that implicit copies are problematic. In your std::string example, having that many String copies in a Java program would be similarly terrible (and this sometimes happens by accident because of abstraction layers that hide all the copying going on under the covers).
I do think Rust gets a lot of stuff right, but Rust's cognitive load is broadly recognized. I tend to see it as C++ with a lot fewer foot guns. ;-)
> Yes, in C/C++ heap allocations are proportionately pretty expensive, but I've seen Java programs have just ridiculous amounts of implicit allocations but there really isn't much of a problem.
Java programs make "ridiculous amounts of implicit allocations" because allocations are cheap in Java. And they need to be cheap because Java doesn't have value semantics so it leans hard on escape analysis + cheap allocations.
I agree with the rest of your comment, although I think most of Rust's "cognitive load" amounts to borrow-checker-vs-garbage-collection. You could envision a Rust with explicit allocations and a GC, and that language would have a "cognitive load" approaching that of Go while also being a fair bit more performant insofar as people can much more easily reason about allocations and thus performance.
> Java programs make "ridiculous amounts of implicit allocations" because allocations are cheap in Java. And they need to be cheap because Java doesn't have value semantics so it leans hard on escape analysis + cheap allocations.
Yes, but that's kind of the point, right? Implicit allocation isn't really a problem because a runtime that optimizes the allocations magically for you is a lot easier to build than a runtime that optimizes whether you really need to be copying objects as much as you do.
> Implicit allocation isn't really a problem because a runtime that optimizes the allocations magically for you is a lot easier to build
As far as I know, Java's (default) runtime gives cheap allocations at the cost of long GC pause times.
> than a runtime that optimizes whether you really need to be copying objects as much as you do
It's not "copying", it's "allocating", and avoiding allocations isn't that much work (and frankly I'm surprised it's such a minor problem that no one has bothered to build an IDE plugin that highlights these allocation points automatically--or at least I haven't heard of such a thing). Anyway, "a runtime that minimizes allocations" is just an escape analyzer and Java has one of these too, and IIRC it's a lot more sophisticated than Go's (but it's also a lot harder to reason about as a consequence).
> As far as I know, Java's (default) runtime gives cheap allocations at the cost of long GC pause times.
"long GC pause times" is kind of vague, so I guess you could be correct, but in practice there's a LOT of different ways the memory management can be handled, many of which are deemed "pauseless GC" (though the term is somewhat misleading).
My statement was considering that reality though. While not true for some use cases, in the vast majority of cases, the runtime optimizes the allocations more than sufficiently.
> It's not "copying", it's "allocating"
Allocators can do a pretty good job of minimizing the overhead of allocation, to the point the amortized cost isn't much more than a single machine instruction. Allocating gigabytes of memory quickly is possible. Copying the data can be a lot more work, and often objects have copy semantics that add a lot more additional work.
> Anyway, "a runtime that minimizes allocations" is just an escape analyzer and Java has one of these too, and IIRC it's a lot more sophisticated than Go's (but it's also a lot harder to reason about as a consequence).
I think you're implicitly saying "a runtime that minimizes heap allocations" there, in which case I'd agree.
> in practice there's a LOT of different ways the memory management can be handled, many of which are deemed "pauseless GC" (though the term is somewhat misleading).
Yes, but I'm pretty sure those "pauseless GC" schemes impose other tradeoffs.
> My statement was considering that reality though. While not true for some use cases, in the vast majority of cases, the runtime optimizes the allocations more than sufficiently.
I'm not sure I follow. The same could be said for Go--in the vast majority of cases, Go's tradeoffs (slow allocations, low latency / non-moving GC) are also suitable.
> Allocators can do a pretty good job of minimizing the overhead of allocation, to the point the amortized cost isn't much more than a single machine instruction.
As far as I know, speeding up allocations to this degree requires a moving GC which imposes a bunch of other constraints (including copying a bunch of memory around).
> Allocating gigabytes of memory quickly is possible. Copying the data can be a lot more work, and often objects have copy semantics that add a lot more additional work.
Yes, but the bottleneck here wasn't the copying, it was the allocations. And if you optimized away allocation cost entirely such that only the copy cost remained, that cost would be so small that the OP would never have bothered to profile because copying small objects like this is so cheap compared to everything else (even if it is expensive compared to bump allocating).
> I think you're implicitly saying "a runtime that minimizes heap allocations" there, in which case I'd agree.
Yes, the allocator and GC are concerned with heap allocations and not stack allocations. I'm using "allocations" as a shorthand for "heap allocations".
In hindsight, I think I chose how to present this poorly, because yes, in this case, the allocation is what is killing the performance. I look at it, and I just see unnecessary implied behaviour creating a performance problem. Usually it isn't the allocations themselves that kill you, but it certainly is the case here.
I agree with you (I think?) that the implicit allocations are a pain point. I think in the Go case, it is the allocations that kill you most of the time (at least that's the case for me), but in C++ it's more likely to be expensive copy constructors or destructors or so on.
Allocations are as damaging as your free function is slow.
Java has a tremendously good GC, so can cope with lots of allocations. Go has an OK one, so needs some help (but mollifying it often pays dividends elsewhere in locality and memory usage too). C++ has your default system heap, good luck.
Historically Java has traded long pause times for fast allocations, although I'm of the impression that it has recently found a way to have its cake and eat it.
Java has been tunable for a long time. Periodically, the recommended tuning changes, or new GC algorithms become available, etc. But it has long been possible to get short pause times with various combinations of choosing the right algorithm and writing your program the right way.
I think what really throws people off here is that getting good performance out of a Java application involves some skills which are alien to C++ programmers, and vice versa. You take an experienced C++ programmer and drop them into a Java codebase, they may have a very poor sense of what is expensive and what is cheap. Vice versa… experienced Java programmers don’t do well in C++ either.
The result is that you have precious few people with any significant, real-world experience fixing performance issues in both languages.
Agreed, but usually tuning for short pause times involves trading off throughput or allocation performance. But at the end of the day, if you aren't allocating a bunch of garbage in the first place, then you don't need to be as concerned about the costs of allocating or cleaning up the garbage. I wish Go did more to make allocations explicit so they could be more easily recognized and avoided; I dislike Java's approach of making allocations even more implicit/idiomatic while trying to fix the cost problem in the runtime (although I admire the ambition).
Parallel garbage collection in Java has been a thing for a long time. With the right tuning you might have very infrequent STW GCs, even in a game as allocation-heavy as Minecraft.
I’ve had no issues with Java 17+ under heavy allocation/garbage collection (data encryption pipeline I haven’t tuned to reuse buffers yet), and it’s pause times are on the order of a handful of milliseconds, without meaningful tuning. I think it’s doing something like a GB/s of garbage collection.
And the jvm in question is doing a LOT more than just this, so it’s coping with millions of allocated objects at the same time.
I consider tens of milliseconds to be a long pause time (P99 should be <10ms), which is what I understand to be the ballpark for G1 (I haven't tested 17+). No doubt JVM is a fine piece of engineering, but I prefer Go's approach of just not allocating as much to begin with (Java's cheap allocations require a moving GC, which introduces challenges with respect to rewriting pointers and so on, which in turn introduces constraints for FFI and makes the whole system more complex). I wish it went further and made allocations more explicit.
Yeah, I’m nominally familiar with these, but I can’t understand why one of these low latency collectors wouldn’t be the default GC unless they impose other significant tradeoffs.
They don’t change it from the default because G1 isn’t terrible and Java has a long history of ensuring backwards compatibility - not just in APIs, but also behavior, warts and all.
If they changed the default GC, a lot of folks would freak out, even if it was objectively better in nearly every situation. Because someone, somewhere is relying on some weird bug in G1, or whatever and now their software behaves differently, and it’s a problem.
Give it a couple more major revs, and it might still happen. G1 became the default in what, Java 9?
Naturally there are tradeoffs, for example the amount of infrastruture they need only justifies when one is working at such scale, for example ZGC requires large pages and makes use of NUMA if available.
I’m not sure I get what you mean. You wouldn’t have that many String copies in Java by passing an unchanged String down the call stack. My point is that it’s too easy to make this mistake in C++.
In Java, the mistake happens only when there's an abstraction that hides the copying from you, so it isn't implicit in the same way, but it's still implicit.
> Rust gets this right, with the hindsight of C++’s example: “a = b” is a move operation by default and clone() is always explicit
Note that a move can still do a copy; in fact, Rust is kinda notorious for generating more on-stack memory copy operations than C++. It’s slowly improving, but it can still be surprisingly bad in some cases.
> except for plain data types where copying is literally memcpy
what do you mean by this? If I say `let x = 5; let y = x;` in rust, that's a "plain data type copy" of a stack value, but memcpy is usually used to copy heap memory. What connection between copying of primitive simple stack values and memcpy are you suggesting here?
The compiler can optimize memcpy with a known size into a small number of move instructions so they are identical to copying stack values.
Try playing with memcpy on Godbolt and you'll find that the compiler will compile the memcpy to a single mov instruction when the size is small, and some movdqu/movups when the size is slightly large, and only a function call when the size is huge.
> memcpy is usually used to copy heap memory
memcpy is often used in low-level serialization / deserialization code since you can't just cast a buffer pointer to a uint32_t pointer and dereference that; the solution is memcpy between variables that are often both on the stack.
> What connection between copying of primitive simple stack values and memcpy are you suggesting here?
They're just using 'memcpy' as a shorthand for saying the bitpattern is blitted. Semantically, that's like a memcpy. The point is, there are no expensive allocations, nor do any embedded pointer fields need adjusted, etc.
Why do you think memcpy is normally used to copy heap memory? It's just a general bitwise copy from one location to the other.
I think the confusion here is that there isn't always a literal call to memcpy for copying small types like ints in the emitted code, but it's always doing something with the same effect and maybe sometimes using an actual memcpy (probably when copying arrays?).
Also something interesting is that memcpy is used for copying data between stack variables in C sometimes when you need to convert some type to another one without using a cast.
I disagree--the bottleneck here is entirely the allocation. The copying is just a memcpy and it's very fast for small structs like this; like I said, it's not the same as a clone() in Rust, which is a deep copy. If you optimized the allocation away entirely (leaving only the copy cost), there wouldn't have been a significant performance problem and this blog would never have been written.
Actually, you'll find that in Rust, Box::new(stuff) will too often put stuff on the stack before copying it in the newly allocated memory. For large enough stuff, that can be slower than the allocation.
> I've only ever found myself missing value semantics when I use languages with implicit reference semantics.
Oh, I miss it every time. ;-)
I will say though that some newer languages seem to have a confused idea about how to offer mixed semantics. A bunch of them tie semantics to types. The ideal interface can vary by usage context. It's hard enough getting the semantics right as the callee (as opposed to caller), let alone when you're defining a type that will be used by who knows how many interfaces.
> I guess I always figured the solution was "value semantics with better education / tooling".
I've always thought much the same, but I have slowly come to appreciate that it's more than just education & tooling. Even with good education & tooling, there's a cognitive load that comes with getting interfaces right that for the general case is just not worth it.
I think this is half right. For anything 64 bits or smaller, value semantics are pretty much always going to be better. That said, being able to choose between value and reference semantics for larger objects per object is a pretty useful feature.
> For anything 64 bits or smaller, value semantics are pretty much always going to be better.
That's assuming a 64-bit CPU (which admittedly seems like a reasonable assumption. The nice thing about the abstraction though is that there's nothing preventing the runtime from applying value semantics for those trivial small-object cases where they're obviously more efficient.
Even for a 32-bit CPU a 64-bit type is only two words to copy - and in many cases those "copies" are just register loads. In contrast, reference types means to even access it you have to read the reference and then indirectly load the memory it points to. You have to really make something contrived where a two-word type ends up being more efficient as a reference than as a value.
Yeah, I never cared for that. Specifically, I'd prefer that everything was just "struct", but structs could implement interfaces, which is essentially the Go/Rust model.
The counter argument to the "explicit is better than implicit" is that abstraction & encapsulation are such significant force multipliers. If done properly, implicit is good. It's just that in case of copying, doing it "properly" is well nigh impossible.
I think the main issue in C++ isn't value semantics, it's deep copy semantics. E.g. in a functional language ADTs are immutable and don't have identity. They can be freely copied, or not, passed by reference or by value, but they are never deep copied. Comparison may be deep, but not passing them.
That is to say, I think I mostly am agreeing with you. In Java, objects are always passed by reference, never by value, and never implicitly copied. But Java doesn't have any value types other than primitives. When I added ADTs to Virgil, I wanted to get the best of both worlds; objects are pass by reference, like Java, and ADTs are immutable, identity-less, so they are deep compared but never deep copied. (ADTs in Virgil can also sometimes be flattened, so they don't necessarily result in heap allocation).
In a functional language, you don't have to worry about bits of code mutating your data. ;-) On the flip side, there's a lot of cognitive load that comes with functional languages, so while they do address the problem neatly...
I'd have to take a look at Virgil to appreciate your approach, but I'm always leery of implicit value vs. reference semantics tied to types (aside from the whole array fiasco, easily the ugliest part of Java's type system). So often the particular semantics you want are driven by the context of what you're doing, rather than the what you're doing it with.
I still don't see why value structs need to be immutable; ints are mutable in all languages, and structs are mutable in C, C++, and Rust (if you `let mut`) and it's a feature of the language.
The performance issue here is not value semantics, it's the overhead of automatic lifetime management. The copy is cheap. The lifetime tracking is not because it forces a heap allocation and creates additional GC pressure. In fact, assuming Rule is small, if Match returned by value, the code would be similarly as fast.
> The performance issue here is not value semantics
There's no performance cost to value semantics, so of course not.
> The copy is cheap. The lifetime tracking is not because it forces a heap allocation and creates additional GC pressure. In fact, assuming Rule is small, if Match returned by value, the code would be similarly as fast.
I'm referring more to how this stuff seeps in without the programmer realizing it. It's the implicit nature of all this behaviour that is the problem.
Sorry, because of value semantics. (Also, you wrote "performance costs of inopportune value semantics". So the correction is annoying; you know what I meant.)
This stuff seeps in without the programmer realizing it because Go made a deliberate design decision to have automatic lifetime management. In other words, this is a feature of the language and not a bug. The only way for this to not seep in is if Go forced programmers to specify most lifetimes, which would make the language much more cumbersome to use.
I.e., this is not a value-vs-reference semantics issue. It's a manual-lifetime-management vs automatic-lifetime-management issue. The solution is to either 1) write in a language with more explicit lifetimes if performance is that important, or 2) profile your code and reduce heap allocations, which is what the person who wrote the article did.
I would disagree about this stuff seeping in because of automatic lifetime management. The equivalent code in Java, Smalltalk, etc., would either not have the performance problem or it would be much more obvious that the code wasn't as efficient as it could be.
> Sorry for being annoying. I wasn't trying to be.
Thanks, no worries.
You're right that Java doesn't have this performance problem because it (apart from primitives) has no implicit copies. (So the code wouldn't copy, it would return a reference, hence no allocation.)
I think what you're trying to say is: programs written languages with value semantics can implicitly create more objects, which (in some cases) have additional allocation/lifetime tracking in languages with automatic lifetime management. So maybe one solution is to prevent copies in most circumstances, and when you do need to copy, make them explicit.
But I think this conflicts with Go's philosophy as well. First, in highly concurrent programs, copies are necessary. A mutex/atomic will become a bottleneck; sharing is bad for performance. This means that copies should be part of most types in the language. (If they aren't, you'll run into issues like not being able to pickle when using multiprocessing in Python [1].)
OK, so we need copies. But clearly, I shouldn't have to write `a = b.Copy()` if b is an integer. That's too verbose. And if everything has a `.Copy()` call, that leads to visual noise and it's not clear when copies are actually expensive. So what set of "blessed" types should be implicitly copyable? Java picks primitives, but this means it's basically impossible to create an int-like object with value semantics. I think this is bad. C++ is at the other extreme with "(almost) all objects are implicitly copyable" but some people dislike how forgetting a & on a std::vector type might lead to a huge copy.
Go has chosen a reasonable middle ground -- dynamically sized types like arrays and maps are not implicitly copyable, but "cheap" objects like structs are implicitly copyable. This means that when you see a Copy call, it's meaningful. But this means you run into these situations when an implicit copy interacting with lifetimes/garbage collection causes a slowdown. But I don't see any other alternative. Getting rid of implicit copies for cheap types will cause visual noise. Getting rid of GC makes your language harder to use. Both of these alternatives seem worse to me.
To be clear, I'm not a huge fan of Go. Especially around the lack of monadic error handling. But I think they have done a decent job wrt. their design goals.
Becoming good at Go is mostly knowing all the sharp edges with the built in types IMO. Of course go routines and concurrency primitives are difficult to master, but that is a different beast and if you understand concurrency from some other languages Go just makes that easy. But knowing all the behaviors of slices, and really intuiting how they work makes your life a lot easier and a lot less prone to bugs. And slice behavior almost all comes down to what is a slice internally and how and where am I copying things as I wrote my code. Generally the copying is fine, but in some cases it is not or it is in a tight performance critical section where you need to be thinking about it.
Esit: Also, pointers into slices will probably leave you sad. You get a pointer to the slice storage, not a pointer to the thing. And if the slice resizes your pointer mow references a dead slice. Basically, pointers and slices are not friends. Unless you have a slice of pointers, which idiomatic go avoids unless there is a decent reason for it :)
I'm a dumb dumb. Can you define "implicit reference semantics" and "value semantics"? You use the phrases several times in your post, but I don't really understand what you mean. If it helps, I'm not a C++ programmer, but I am familiar with higher level languages like Go, Python, Ruby, PHP and Javascript.
It's the difference between assigning/passing around "copies of the data" vs. assigning/passing around "the memory address for that data" under the hood.
PHP, for example, has explicit references. If you have an `$arr1=array(1,2,3)` and an `$arr2 = $arr1`, that second array is a full copy of the first array, and updating $arr1 does nothing to $arr2. Similarly, `function update_array($arr) { $arr[0] = 'cake'; }` called with `$arr1` will create a copy of $arr1 for use inside that function. Any changes to `$arr` inside the function only apply to that copy, not to $arr1. Unless you explicitly tell PHP you want to work with references, by using `function update_array(&$arr) { ... }` instead. PHP uses value semantics.
JS, on the other hand, uses implicit reference semantics. If you have a `const arr1 = [1,2,3];` then `arr1` is an alias, simply pointing to a memory location, and declaring a `const arr2 = arr1`, makes arr2 also be an alias for that same memory location. They both reference the same thing, but neither "is" the thing. Similarly, if you have a `function updateArray(arr) { arr[0] = 'cake'; }` then that `arr` is also just an alias pointing to the memory location of whatever you passed in, and changes to `arr` change the underlying data, so whether you try to access that through `arr1`, `arr2` and `arr`, it's the same thing. JS uses implicit reference semantics.
(But note that many languages with implicit reference semantics make exceptions for immutable primitives like numbers or strings)
That's not technically correct with regards to PHP. Your statement that any changes to $arr1 or $arr2 only impact the one in question, however, is accurate. If no changes are made they still refer to the same data in memory. It's copy-on-write semantics.
$arr1 = [1,2,3];
// $arr1 is a pointer to a zval array [1,2,3] and refcount:1
$arr2 = $arr1;
// $arr1 and $arr2 are pointers to the same zval array but incremented refcount to 2.
$arr2[] = 4;
// at this point, $arr2 creates a new zval array, copies over the data from the first, sets the recount to 1, and appends int(4). The [1,2,3] array has its refcount decremented to 1 as it's still referenced by $arr1.
Actually that is outdated since at least PHP 7. In modern PHP the engine uses copy on write only for "larger" things, but "small" things like integers or booleans live on the stack and are copied around, avoiding heap allocations etc.
This isn't semantics, though, but implementation. As far as I'm aware there's no way to "observe" the copy short of digging into runtime debugging tools, so I think it's still fair to say the language has pass-by-value semantics. In C++ we still refer to things returned-by-value as returned by value despite the reality of NRVO, etc.
“Implicit reference semantics” means that variables ordinarily refer to objects rather than containing them. “Value semantics” means that variables contain values rather than references, though there are pointer types that let you explicitly store references when you want them. (Often this is discussed in terms of parameter passing, for historical reasons, but the same ideas apply to variables more generally.)
If your language has implicit reference semantics, “x = y” will cause x and y to refer to the same object. If it has value semantics, x will be a copy of y.
"Implicit reference semantics" = everything (or near enough) is implicitly a reference. That's what you get in Python, Ruby, or Javascript: when you pass something around, you're really passing a reference to it, a pointer. Everyone ends up referring to the same thing, and sharing it.
"Value semantics": you pass the value itself around, which means you're shallow-copying it all the time. That's what you get when you pass or return a bare non-interface type in Go.
PHP is a bit of a mess, objects are passed by reference (= implicit reference semantics) but arrays are passed by value, and you can opt them into pass-by-reference. You can also pass non-arrays by reference, though I don't think that's very common.
Languages like Python, Java, and C# have implicit reference semantics. When you create an object, you get a pointer to that object (usually, or commonly).
In languages like C, C++, Go, Rust… references are more explicit. If you want a pointer to an object, you have to &, or something similar.
If you're familiar with Python and Go, you'll likely be able to quickly spot the differences in how they handle parameter passing. Python uses references and Go uses value.
Go maps and channels have reference semantics. Slices are passed and returned by value, but backing arrays are often shared (not copied) in an error-prone way, so it’s safest to pretend they have move semantics and treat only one copy as valid.
IMHO they should have used pointers everywhere they didn’t want value semantics with deep copies.
Having used C++ relatively little, and a long time ago (and never used Go), I don't think I ever realized that value semantics could be used to implicitly clone heap objects
Normally they’re not, that’s really specific to C++ nonsense.
In a normal value-semantics system if you have a pointer you just copy the pointer. Obviously if the langage doesn’t have your back it also means you’ve now fucked up your ownership, but you’re in C so that was to be expected.
Yeah, this was my assumption when I read the original comment, so I was confused about what it was saying (I thought it was suggesting you should pass references instead of doing a bitwise-copy most of the time, which is just plausible enough that I believed it for a minute) until I read further down in the thread and realized they meant implicit heap-cloning
I would also lay a chunk of blame on the use of type inference which causes the information about the behaviour to be hidden from view.
Had the author been looking at the type information within the syntax of the code the profile output may not have been a surprise. Perhaps the problem would never have existed in the first place.
Yeah. Often, writing out the type explicitly is just busywork. But it seems it would have paid off here.
If you were forced to stop and think what type to declare, I bet you'd write "var rule *Rule". Even if you don't think deeply and just look at the return type.
And then if you assigned "r[i]" to "rule", you'd get a type error.
It's not that value semantics can't be better (they most assuredly can be), or that reference semantics don't cause their own complexity problems, but rather that so often we thoughtlessly imply/impose value semantics through interfaces in ways that negatively impact performance; getting interfaces wrong is a much tougher bell to unring.
The vast majority of my mental energy when I define an interface in C++ is carefully thinking through a combination of ownership contracts and value vs. reference semantics that I can mostly ignore in languages with implicit reference semantics. While occasionally ignoring those contracts while developing in Java/Python/whatever comes back to bite me, the problem isn't nearly as common or problematic as when I unintentionally impose value semantics in a language that allows me to.