2019-04-04 14:58:14 +02:00

16 KiB

Raw Blame History

Efficient Params and String Formatting

Summary

This combination of features will increase the efficiency of formatting string values and passing of params style arguments.

Motivation

The allocation overhead of formatting string values can dominate the performance of many text based applications: from the boxing penalty of struct types, the object[] allocation for params and the intermediate string allocations during string.Format calls. In order to maintain efficiency such applications often need to abandon productivity features such as params and string interpolation and move to non-standard, hand coded solutions.

Consider MSBuild as an example. This is written using a lot of modern C# features by developers who are conscious of performance. Yet in one representative build sample MSBuild will generate 262MB of string allocation using minimal verbosity. Of that 1/2 of the allocations are short lived allocations inside string.Format. These features would remove much of that on .NET Desktop and get it down to nearly zero on .NET Core due to the availability of Span<T>

The set of language features described here will enable applications to continue using these features, with very little or no churn to their application code base, while removing the unintended allocation overhead in the majority of cases.

Detailed Design

There are a set of features that will be used here to achieve these results:

Expanding params to support a broader set of collection types.
Allowing for developers to customize how string interpolation is achieved.
Allowing for interpolated string to bind to more efficient string.Format overloads.

Extending params

The language will allow for params in a method signature to have the types Span<T>, ReadOnlySpan<T> and IEnumerable<T>. The same rules for invocation will apply to these new types that apply to params T[]:

Can't overload where the only difference is a params keyword.
Can invoke by passing a series of arguments that are implicitly convertible to T or a single Span<T> / ReadOnlySpan<T> / IEnumerable<T> argument.
Must be the last parameter in a method signature.
Etc ...

The Span<T> and ReadOnlySpan<T> variants will be referred to as Span<T> below for simplicity. In cases where the behavior of ReadOnlySpan<T> differs it will be explicitly called out.

The advantage the Span<T> variants of params provides is it gives the compiler great flexibility in how it allocates the backing storage for the Span<T> value. With a params T[] the compiler must allocate a new T[] for every invocation of a params method. Re-use is not possible because it must assume the callee stored and reused the parameter. This can lead to a large inefficiency in methods with lots of params invocations.

Given Span<T> variants are ref struct the callee cannot store the argument. Hence the compiler can optimize the call sites by taking actions like re-using the argument. This can make repeated invocations very efficient as compared to T[]. The language though will make no specific guarantees about how such callsites are optimized. Only note that the compiler is free to use values other than T[] when invoking a params Span<T> method.

One such potential implementation is the following. Consider all params invocation in a method body. The compiler could allocate an array which has a size equal to the largest params invocation and use that for all of the invocations by creating appropriately sized Span<T> instances over the array. For example:

static class OneAllocation {
    static void Use(params Span<string> spans) {
        ...
    }

    static void Go() {
        Use("jaredpar");
        Use("hello", "world");
        Use("a", "longer", "set");
    }
}

The compiler could choose to emit the body of Go as follows:

    static void Go() {
        var args = new string[3];
        args[0] = "jaredpar";
        Use(new Span<string>(args, start: 0, length: 1));

        args[0] = "hello";
        args[1] = "world";
        Use(new Span<string>(args, start: 0, length: 2));

        args[0] = "a";
        args[1] = "longer";
        args[2] = "set";
        Use(new Span<string>(args, start: 0, length: 3));
   }

This can significantly reduce the number of arrays allocated in an application. Allocations can be even further reduced if the runtime provides utilities for smarter stack allocation of arrays.

This optimization cannot always be applied though. Even though the callee cannot capture the params argument it can still be captured in the caller when there is a ref or a out / ref parameter that is itself a ref struct type.

static class SneakyCapture {
    static ref int M(params Span<T> span) => ref span[0];

    static void Oops() {
        // This now holds onto the memory backing the Span<T> 
        ref int r = ref M(42);
    }
}

These cases are statically detectable though. It potentially occurs whenever there is a ref return or a ref struct parameter passed by out or ref. In such a case the compiler must allocate a fresh T[] for every invocation.

Several other potential optimization strategies are discussed at the end of this document.

The IEnumerable<T> variant is a merely a convenience overload. It's useful in scenarios which have frequent uses of IEnumerable<T> but also have lots of params usage. When invoked in T argument form the backing storage will be allocated as a T[] just as params T[] is done today.

params overload resolution changes

This proposal means the language now has four variants of params where before it had one. It is sensible for methods to define overloads of methods that differ only on the type of a params declarations.

Consider that StringBuilder.AppendFormat would certainly add a params ReadOnlySpan<object> overload in addition to the params object[]. This would allow it to substantially improve performance by reducing collection allocations without requiring any changes to the calling code.

To facilitate this the language will introduce the following overload resolution tie breaking rule. When the candidate methods differ only by the params parameter then the candidates will be preferred in the following order:

ReadOnlySpan<T>
Span<T>
T[]
IEnumerable<T>

This order is the most to the least efficient for the general case.

Variant

CoreFX is prototyping a new managed type named Variant. This type is meant to be used in APIs which expect heterogeneous values but don't want the overhead brought on by using object as the parameter. The Variant type provides universal storage but avoids the boxing allocation for the most commonly used types. Using this type in APIs like string.Format can eliminate the boxing overhead in the majority of cases.

This type itself is not necessarily special to the language. It is being introduced in this document separately though as it becomes an implementation detail of other parts of the proposal.

Efficient interpolated strings

Interpolated strings are a popular yet inefficient feature in C#. The most common syntax, using an interpolated string as a string, translates into a string.Format(string, params object[]) call. That will incur boxing allocations for all value types, intermediate string allocations as the implementation largely uses object.ToString for formatting as well as array allocations once the number of arguments exceeds the amount of parameters on the "fast" overloads of string.Format.

The language will change its interpolation lowering to consider alternate overloads of string.Format. It will consider all forms of string.Format(string, params) and pick the "best" overload which satisfies the argument types. The "best" params overload will be determined by the rules discussed above. This means interpolated string can now bind to very efficient overloads like string.Format(string format, params ReadOnlySpan<Variant> args). In many cases this will remove all intermediate allocations.

Customizable interpolated strings

Developers are able to customize the behavior of interpolated strings with FormattableString. This contains the data which goes into an interpolated string: the format string and the arguments as an array. This though still has the boxing and argument array allocation as well as the allocation for FormattableString (it's an abstract class). Hence it's of little use to applications which are allocation heavy in string formatting.

To make interpolated string formatting efficient the language will recognize a new type: System.ValueFormattableString. All interpolated strings will have a target type conversion to this type. This will be implemented by translating the interpolated string into the call ValueFormattableString.Create exactly as is done for FormattableString.Create today. The language will support all params options described in this document when looking for the most suitable ValueFormattableString.Create method.

readonly struct ValueFormattableString {
    public static ValueFormattableString Create(Variant v) { ... } 
    public static ValueFormattableString Create(string s) { ... } 
    public static ValueFormattableString Create(string s, params ReadOnlySpan<Variant> collection) { ... } 
}

class ConsoleEx { 
    static void Write(ValueFormattableString f) { ... }
}

class Program { 
    static void Main() { 
        ConsoleEx.Write(42);
        ConsoleEx.Write($"hello {DateTime.UtcNow}");

        // Translates into 
        ConsoleEx.Write(ValueFormattableString.Create((Variant)42));
        ConsoleEx.Write(ValueFormattableString.Create(
            "hello {0}", 
            new Variant(DateTime.UtcNow));
    }
}

Overload resolution rules will be changed to prefer ValueFormattableString over string when the argument is an interpolated string. This means it will be valuable to have overloads which differ only on string and ValueFormattableString. Such an overload today with FormattableString is not valuable as the compiler will always prefer the string version (unless the developer uses an explicit cast).

Open Issues

ValueFormattableString breaking change

The change to prefer ValueFormattableString during overload resolution over string is a breaking change. It is possible for a developer to have defined a type called ValueFormattableString today and use it in method overloads with string. This proposed change would cause the compiler to pick a different overload once this set of features was implemented.

The possibility of this seems reasonably low. The type would need the full name System.ValueFormattableString and it would need to have static methods named Create. Given that developers are strongly discouraged from defining any type in the System namespace this break seems like a reasonable compromise.

Expanding to more types

Given we're in the area we should consider adding IList<T>, ICollection<T> and IReadOnlyList<T> to the set of collections for which params is supported. In terms of implementation it will cost a small amount over the other work here.

LDM needs to decide if the complication to the language is worth it though. The addition of IEnumerable<T> removes a very specific friction point. Lacking this params solution many customers were forced to allocate T[] from an IEnumerable<T> when calling a params method. The addition of IEnumerable<T> fixes this though. There is no specific friction point that the other interfaces fix here.

Considerations

Variant2 and Variant3

The CoreFX team also has a non-allocating set of storage types for up to three Variant arguments. These are a single Variant, Variant2 and Variant3. All have a pair of methods for getting an allocation free Span<Variant> off of them: CreateSpan and KeepAlive. This means for a params Span<Variant> of up to three arguments the call site can be entirely allocation free.

static class ZeroAllocation {
    static void Use(params Span<Variant> spans) {
        ...
    }

    static void Go() {
        Use("hello", "world");
    }
}

The Go method can be lowered to the following:

static class ZeroAllocation {
    static void Go() {
        Variant2 _v;
        _v.Variant1 = new Variant("hello");
        _v.Variant2 = new Variant("word");
        Use(_v.CreateSpan());
        _v.KeepAlive();
    }
}

This requires very little work on top of the proposal to re-use T[] between params Span<T> calls. The compiler already needs to manage a temporary per call and do clean up work after (even if in one case it's just marking an internal temp as free).

Note: the KeepAlive function is only necessary on desktop. On .NET Core the method will not be available and hence the compiler won't emit a call to it.

CLR stack allocation helpers

The CLR only provides only localloc for stack allocation of contiguous memory. This instruction is limited in that it only works for unmanaged types. This means it can't be used as a universal solution for efficiently allocating the backing storage for params Span<T>.

This limitation is not some fundamental restriction though but instead more an artifact of history. The CLR could choose to add new op codes / intrinsics which provide universal stack allocation. These could then be used to allocate the backing storage for most params Span<T> calls.

static class BetterAllocation {
    static void Use(params Span<string> spans) {
        ...
    }

    static void Go() {
        Use("hello", "world");
    }
}

The Go method can be lowered to the following:

static class ZeroAllocation {
    static void Go() {
        Span<T> span = RuntimeIntrinsic.StackAlloc<string>(length: 2);
        span[0] = "hello";
        span[1] = "world";
        Use(span);
    }
}

While this approach is very heap efficient it does cause extra stack usage. In an algorithm which has a deep stack and lots of params usage it's possible this could cause a StackOverflowException to be generated where a simple T[] allocation would succeed.

Unfortunately C# is not set up for the type of inter-method analysis where it could make an educated determination of whether or not call should use stack or heap allocation of params. It can only really consider each call on its own.

The CLR is best setup for making this type of determination at runtime. Hence we'd likely have the runtime provide two methods for universal stack allocation:

Span<T> StackAlloc<T>(int length): this has the same behaviors and limitations of localloc except it can work on any type T.
Span<T> MaybeStackAlloc<T>(int length): this runtime can choose to implement this by doing a stack or heap allocation. The runtime can then use the execution context in which it's called to determine how the Span<T> is allocated. The caller though will always treat it as if it were stack allocated.

For very simple cases, like one to two arguments, the C# compiler could always use StackAlloc<T> variant. This is unlikely to significantly contribute to stack exhaustion in most cases. For other cases the compiler could choose to use MaybeStackAlloc<T> instead and let the runtime make the call.

How we choose will likely require a deeper investigation and examination of real world apps. But if these new intrinsics are available then it will give us this type of flexibility.

Why not varargs?

The existing varargs feature was considered here as a possible solution. This feature though is meant primarily for C++/CLI scenarios and has known holes for other scenarios. Additionally there is significant cost in porting this to Unix. Hence it wasn't seen as a viable solution.

This spec is related to the following issues:

16 KiB Raw Blame History