This is my third post in the GC series. In the first post, I introduced D’s garbage collector and the language features that require it, and touched on simple strategies to use it effectively. In the second post, I showed off the tools provided by the language and library to disable or prohibit the GC in specific parts of a code base, how to use the compiler to assist in that endeavor, and recommended that D programs be written initially to embrace the GC, taking advantage of simple strategies to mitigate its impact, and later tuned to avoid it or further optimize its usage only when profiling shows it’s warranted.

When garbage collection is turned off via GC.disable or prevented by the @nogc function annotation, memory will still need to be allocated from somewhere. And even when the GC is fully embraced, it’s still desirable to minimize the size and quantity of GC heap allocations. That means allocating either via the stack, or via the non-GC heap. The focus of this post is the former. Non-GC heap allocations will be covered in my next post in this series.

Stack allocation

The simplest allocation strategy in D is the same as it is in C: avoid the heap and use the stack whenever possible. When a local array is needed and the size can be known at compile time, use a static rather than a dynamic array. Structs, which are value types and stack-allocated by default, should be preferred where possible over classes, which are reference types and are usually allocated from one heap or another. D’s compile-time features can present opportunities here that might not otherwise be available.

Static arrays

Static array declarations in D require the length to be known at compile-time.

// OK
int[10] nums;

// Error: variable x cannot be read at compile time
int x = 10;
int[x] err;

Unlike dynamic arrays, static arrays can be initialized with array literals with no allocation taking place on the GC heap. The lengths must match, otherwise the compiler will emit an error.

@nogc void main() {
    int[3] nums = [1, 2, 3];
}

Static arrays are automatically sliced when passed to any function taking a slice as a parameter, making them interchangeable with dynamic arrays.

void printNums(int[] nums) {
    import std.stdio : writeln;
    writeln(nums);
}

void main() {
    int[]  dnums = [0, 1, 2];
    int[3] snums = [0, 1, 2];
    printNums(dnums);
    printNums(snums);
}

When compiling with -vgc to find the potential GC allocations in a program and eliminate them where possible, this is an easy win. Just be wary of situations like the following:

int[] foo() {
    auto nums = [0, 1, 2];

    // Do work with nums...

    return nums;
}

Converting nums in this example to a static array would be a mistake. The return statement in that case would be returning a slice to stack-allocated memory, which is a programming error. Luckily, doing so will generate a compiler error.

On the other hand, if the return is conditional, it may be desirable to heap-allocate the array only when absolutely necessary rather than every time the function is called. In that scenario, a static array can be declared locally and a dynamic copy made on return. Enter the .dup property:

int[] foo() {
    int[3] nums = [0, 1, 2];
    
    // Let x = the result of some work with nums
    bool condtion = x;

    if(condition) return nums.dup;
    else return [];
}

This function still uses the GC via .dup, but only allocates if it needs to and avoids allocation when it doesn’t. Note that [] is equivalent to null in this case, a slice (or dynamic array) with a .length of 0 and a .ptr of null.

Structs vs. classes

Struct instances in D are allocated on the stack by default, but can be allocated on the heap when desired. Stack-allocated structs have deterministic destruction, with their destructors called as soon as the enclosing scope exits.

struct Foo {
    int x;
    ~this() {
        import std.stdio;
        writefln("#%s says bye!", x);
    }
}
void main() {
    Foo f1 = Foo(1);
    Foo f2 = Foo(2);
    Foo f3 = Foo(3);
}

As expected, this prints:

#3 says bye!
#2 says bye!
#1 says bye!

Classes, being reference types, are almost always allocated on the heap. Usually, that’s the GC heap via new, though it could also be the non-GC heap through a custom allocator. But there’s no rule saying they can’t be allocated on the stack. The standard library template std.typecons.scoped allows us to easily do so.

class Foo {
    int x;

    this(int x) { 
        this.x = x; 
    }
    
    ~this() {
        import std.stdio;
        writefln("#%s says bye!", x);
    }
}
void main() {
    import std.typecons : scoped;
    auto f1 = scoped!Foo(1);
    auto f2 = scoped!Foo(2);
    auto f3 = scoped!Foo(3);
}

Functionally, this is identical to the struct example above; it prints the same results. Deterministic destruction is achieved via the core.object.destroy function, which allows destructors to be called outside of GC collections.

Note that neither scoped nor destroy are currently usable in @nogc functions. This isn’t necessarily a problem, as a function doesn’t have to be annotated such to avoid the GC, but it can be a headache if you are trying to fit everything into a @nogc call tree. In future posts, we’ll look at some of the design issues that crop up when using @nogc and how to avoid them.

Generally, when implementing custom types in D, the choice between struct and class should be dependent on the type’s intended usage. POD types are obvious candidates for struct, whereas for types in something like a GUI system, where inheritance heirarchies and runtime interfaces are extremely useful, class is a more appropriate choice. Beyond those obvious cases, there are a number of other considerations which could be the focus of a separate blog post on the topic. For our purposes, just keep in mind that whether or not a type is implemented as a struct or a class need not always dictate whether or not instances can be allocated on the stack.

alloca

Given that D makes alloca available, it is also an option for stack allocation. This is a candidate especially for arrays when you want to avoid or eliminate a local GC allocation, but the array size is only known at run time. The following example allocates a dynamic array with a runtime size on the stack.

import core.stdc.stdlib : alloca;

void main() {
    size_t size = 10;
    void* mem = alloca(size);

    // Slice the memory block
    int[] arr = cast(int[])mem[0 .. size];
}

The same caution about using alloca in C applies here: be careful not to blow up the stack. And as with local static arrays, don’t return a slice of arr. Return arr.dup instead.

A simple example

Consider an implementation of a Queue data type. An idiomatic implementation in D is going to be a struct that’s templated on the type it’s intended to contain. In Java, collection usage is interface heavy and it’s recommended to declare an instance using an interface type rather than the implementation type. Structs in D can’t implement interfaces, but in many cases they can still be used to program to interfaces thanks to Design by Introspection (DbI). This allows programming to a common interface that is verified via compile-time introspection without the need for an interface type, so it can work with structs, classes and, thanks to Universal Function Call Syntax (UFCS), even free functions (when the functions are in scope).

D’s arrays are an obvious choice as the backing store for a Queue implementation. Moreover, there’s an opportunity to make the backing store a static array when a queue is intended to be bounded with a fixed size. Since it’s already a templated type, an additional parameter, a template value parameter with a default value can easily be added to decide at compile time if the array should be static or not and, if so, how much space it should require.

// A default Size of 0 means to use a dynamic array for the
// backing store; non-zero indicates a static array.
struct Queue(T, size_t Size = 0) 
{
    // This constant will be inferred as a boolean. By making it
    // public, a DbI template outside of this module can determine
    // whether or not the Queue might grow. 
    enum isFixedSize = Size > 0;

    void enqueue(T item) 
    {
        static if(isFixedSize) {
            assert(_itemCount < _items.length);
        }
        else {
            ensureCapacity();
        }
        push(item);
    }

    T dequeue() {
        assert(_itemCount != 0);
        static if(isFixedSize) {
            return pop();
        }
        else {
            auto ret = pop();
            ensurePacked();
            return ret;
        }
    }

    // Only available on a growable array
    static if(!isFixedSize) {
        void reserve(size_t capacity) { /* Allocate space for new items */ }
    }

private:   
    static if(isFixedSize) {
        T[Size] _items;     
    }
    else T[] _items;
    size_t _head, _tail;
    size_t _itemCount;

    void push(T item) { 
        /* Add item, update _head and _tail */
        static if(isFixedSize) { ... }
        else { ... }
    }

    T pop() { 
        /* Remove item, update _head and _tail */ 
        static if(isFixedSize) { ... }
        else { ... }
    }

    // These are only available on a growable array
    static if(!isFixedSize) {
        void ensureCapacity() { /* Alloc memory if needed */ }
        void ensurePacked() { /* Shrink the array if needed */}
    }
}

With this, the client can declare instances like so:

Queue!Foo qUnbounded;
Queue!(Foo, 128) qBounded;

qBounded requires no heap allocations. What happens with qUnbounded depends on the implementation. Moreover, compile-time introspection can be used to test if an instance is a fixed size or not. The isFixedSize constant is a convenience for that. Clients could alternatively use the built-in __traits(hasMember, T, "reserve") or the standard library function std.traits.hasMember!T("reserve") in one compile-time construct or another (__traits and std.traits are great for DbI, and the latter should be preferred when it provides similar functionality), but including the constant in the type is more convenient.

void doSomethingWithQueueInterface(T)(T queue)
{
    static if(T.isFixedSize) { ... }
    else { ... }
}

Conclusion

This has been a brief overview of a few options for stack allocation in D to avoid allocations from the GC heap. Making use of them when possible is an easy way to minimize the size and quantity of GC allocations, a proactive strategy for mitigating potential negative performance impacts from garbage collection.

The next post in this series will cover some of the options available for non-GC heap allocations.

The first post I wrote in the GC series introduced the D garbage collector and the language features that use it. Two key points that I tried to get across in the article were:

1. The GC can only run when memory allocations are requested. Contrary to popular misconception, the D GC isn’t generally going to decide to pause your Minecraft clone in the middle of the hot path. It will only run when memory from the GC heap is requested, and then only if it needs to.
2. Simple C and C++ allocation strategies can mitigate GC pressure. Don’t allocate memory in inner loops – preallocate as much as possible, or fetch it from the stack instead. Minimize the total number of heap allocations. These strategies work because of point #1 above. The programmer can dictate when it is possible for a collection to occur simply by being smart about when GC heap allocations are made.

The strategies in point #2 are fine for code that a programmer writes herself, but they aren’t going to help at all with third-party libraries. For those situations, D provides built-in mechanisms to guarantee that no GC allocations can occur, both in the language and the runtime. There are also command-line options that can help make sure the GC stays out of the way.

Let’s imagine a hypothetical programmer named J.P. who, for reasons he considers valid, has decided he would like to avoid garbage collection completely in his D program. He has two immediate options.

The GC chill pill

One option is to make a call to GC.disable when the program is starting up. This doesn’t stop allocations, but puts a hold on collections. That means all collections, including any that may result from allocations in other threads.

void main() {
    import core.memory;
    import std.stdio;
    GC.disable;
    writeln("Goodbye, GC!");
}

Output:

Goodbye, GC!

This has the benefit that all language features making use of the GC heap will still work as expected. But, considering that allocations are still going without any cleanup, when you do the math you’ll realize this might be problematic. If allocations start to get out of hand, something’s gotta give. From the documentation:

Collections may continue to occur in instances where the implementation deems necessary for correct program behavior, such as during an out of memory condition.

Depending on J.P.’s perspective, this might not be a good thing. But if this constraint is acceptable, there are some additional steps that can help keep things under control. J.P. can make calls to GC.enable or GC.collect as necessary. This provides greater control over collection cycles than the simple C and C++ allocation strategies.

The GC wall

When the GC is simply intolerable, J.P. can turn to the @nogc attribute. Slap it at the front of the main function and thou shalt suffer no collections.

@nogc
void main() { ... }

This is the ultimate GC mitigation strategy. @nogc applied to main will guarantee that the garbage collector will never run anywhere further along the callstack. No more caveats about collecting “where the implementation deems necessary”.

At first blush, this may appear to be a much better option than GC.disable. Let’s try it out.

@nogc
void main() {
    import std.stdio;
    writeln("GC be gone!");
}

This time, we aren’t going to get past compilation:

Error: @nogc function 'D main' cannot call non-@nogc function 'std.stdio.writeln!string.writeln'

What makes @nogc tick is the compiler’s ability to enforce it. It’s a very blunt approach. If a function is annotated with @nogc, then any function called from inside it must also be annotated with @nogc. As may be obvious, writeln is not.

That’s not all:

@nogc 
void main() {
    auto ints = new int[](100);
}

The compiler isn’t going to let you get away with that one either.

Error: cannot use 'new' in @nogc function 'D main'

Any language feature that allocates from the GC heap is out of reach inside a function marked @nogc (refer to the first post in this series for an overview of those features). It’s turtles all the way down. The big benefit here is that it guarantees that third-party code can’t use those features either, so can’t be allocating GC memory behind your back. Another downside is that any third-party library that is not @nogc aware is not going to be available in your program.

Using this approach requires a number of workarounds to make up for non-@nogc language features and library functions, including several in the standard library. Some are trivial, some are not, and others can’t be worked around at all (we’ll dive into the details in a future post). One example that might not be obvious is throwing an exception. The idiomatic way is:

throw new Exception("Blah");

Because of the new in that line, this isn’t possible in @nogc functions. Getting around this requires preallocating any exceptions that will be thrown, which in turn runs into the issue that any exception memory allocated from the regular heap still needs to be deallocated, which leads to ideas of reference counting or stack allocation… In other words, it’s a big can of worms. There’s currently a D Improvement Proposal from Walter Bright intended to stuff all the worms back into the can by making throw new Exception work without the GC when it needs to.

It’s not an insurmountable task to get around the limitations of @nogc main, it just requires a good bit of motivation and dedication.

One more thing to note about @nogc main is that it doesn’t banish the GC from the program completely. D has support for static constructors and destructors. The former are executed by the runtime before entering main and the latter upon exiting. If any of these exist in the program and are not annotated with @nogc, then GC allocations and collections can technically be present in the program. Still, @nogc applied to main means there won’t be any collections running once main is entered, so it’s effectively the same as having no GC at all.

Working it out

Here’s where I’m going to offer an opinion. There’s a wide range of programs that can be written in D without disabling or cutting the GC off completely. The simple strategies of minimizing GC allocations and keeping them out of the hot path will get a lot of mileage and should be preferred. It can’t be repeated enough given how often it’s misunderstood: D’s GC will only have a chance to run when the programmer allocates GC memory and it will only run if it needs to. Use that knowledge to your advantage by keeping the allocations small, infrequent, and isolated outside your inner loops.

For those programs where more control is actually needed, it probably isn’t going to be necessary to avoid the GC entirely. Judicious use of @nogc and/or the core.memory.GC API can often serve to avoid any performance issues that may arise. Don’t put @nogc on main, put it on the functions where you really want to disallow GC allocations. Don’t call GC.disable at the beginning of the program. Call it instead before entering a critical path, then call GC.enable when leaving that path. Force collections at strategic points, such as between game levels, with GC.collect.

As with any performance tuning strategy in software development, it pays to understand as fully as possible what’s actually happening under the hood. Adding calls to the core.memory.GC API in places where you think they make sense could potentially make the GC do needless work, or have no impact at all. Better understanding can be achieved with a little help from the toolchain.

The DRuntime GC option --DRT-gcopt=profile:1 can be passed to a compiled program (not to the compiler!) for some tune-up assistance. This will report some useful GC profiling data, such as the total number of collections and the total collection time.

To demonstrate, gcstat.d appends twenty values to a dynamic array of integers.

void main() {
    import std.stdio;
    int[] ints;
    foreach(i; 0 .. 20) {
        ints ~= i;
    }
    writeln(ints);
}

Compiling and running with the GC profile switch:

dmd gcstat.d
gcstat --DRT-gcopt=profile:1
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]
        Number of collections:  1
        Total GC prep time:  0 milliseconds
        Total mark time:  0 milliseconds
        Total sweep time:  0 milliseconds
        Total page recovery time:  0 milliseconds
        Max Pause Time:  0 milliseconds
        Grand total GC time:  0 milliseconds
GC summary:    1 MB,    1 GC    0 ms, Pauses    0 ms <    0 ms

This reports one collection, which almost certainly happened as the program was shutting down. The runtime terminates the GC as it exits which, in the current implementation, will generally trigger a collection. This is done primarily to run destructors on collected objects, even though D does not require destructors of GC-allocated objects to ever be run (a topic for a future post).

DMD supports a command-line option, -vgc, that will display every GC allocation in a program, including those that are hidden behind language features like the array append operator.

To demonstrate, take a look at inner.d:

void printInts(int[] delegate() dg)
{
    import std.stdio;
    foreach(i; dg()) writeln(i);
} 

void main() {
    int[] ints;
    auto makeInts() {
        foreach(i; 0 .. 20) {
            ints ~= i;
        }
        return ints;
    }

    printInts(&makeInts);
}

Here, makeInts is an inner function. A pointer to a non-static inner function is not a function pointer, but a delegate (a context pointer/function pointer pair; if an inner function is static, a pointer of type function is produced instead). In this particular case, the delegate makes use of a variable in its parent scope. Here’s the output of compiling with -vgc:

dmd -vgc inner.d
inner.d(11): vgc: operator ~= may cause GC allocation
inner.d(7): vgc: using closure causes GC allocation

What we’re seeing here is that memory needs to be allocated so that the delegate can carry the state of ints, making it a closure (which is not itself a type – the type is still delegate). Move the declaration of ints inside the scope of makeInts and recompile. You’ll find that the closure allocation goes away. A better option is to change the declaration of printInts to look like this:

void printInts(scope int[] delegate() dg)

Adding scope to any function parameter ensures that any references in the parameter cannot be escaped. In other words, it now becomes impossible to do something like assign dg to a global variable, or return it from the function. The effect is that there is no longer a need to create a closure, so there will be no allocation. See the documentation for more on function pointers, delegates and closures, and function parameter storage classes.

The gist

Given that the D GC is very different from those in languages like Java and C#, it’s certain to have different performance characteristics. Moreover, D programs tend to produce far less garbage than those written in a language like Java, where almost everything is a reference type. It helps to understand this when embarking on a D project for the first time. The strategies an experienced Java programmer uses to mitigate the impact of collections aren’t likley to apply here.

While there is certainly a class of software in which no GC pauses are ever acceptable, that is an arguably small set. Most D projects can, and should, start out with the simple mitigation strategies from point #2 at the top of this article, then adapt the code to use @nogc or core.memory.GC as and when performance dictates. The command-line options demonstrated here can help ferret out the areas where that may be necessary.

As time goes by, it’s going to become easier to micromanage garbage collection in D programs. There’s a concerted effort underway to make Phobos, D’s standard library, as @nogc-friendly as possible. Language improvements such as Walter’s proposal to modify how exceptions are allocated should speed that work considerably.

Future posts in this series will look at how to allocate memory outside of the GC heap and use it alongside GC allocations in the same program, how to compensate for disabled language features in @nogc code, strategies for handling the interaction of the GC with object destructors, and more.

Thanks to Vladimir Panteleev, Guillaume Piolat, and Steven Schveighoffer for their valuable feedback on drafts of this article.

The article has been amended to remove a misleading line about Java and C#, and to add some information about multiple threads.