Improved Interpolated Strings

Summary

We introduce a new pattern for creating and using interpolated string expressions to allow for efficient formatting and use in both general string scenarios and more specialized scenarios such as logging frameworks, without incurring unnecessary allocations from formatting the string in the framework.

Motivation

Today, string interpolation mainly lowers down to a call to string.Format. This, while general purpose, can be inefficient for a number of reasons:

  1. It boxes any struct arguments, unless the runtime has happened to introduce an overload of string.Format that takes exactly the correct types of arguments in exactly the correct order.
    • This ordering is why the runtime is hesitant to introduce generic versions of the method, as it would lead to combinatoric explosion of generic instantiations of a very common method.
  2. It has to allocate an array for the arguments in most cases.
  3. There is no opportunity to avoid instanciating the instance if it's not needed. Logging frameworks, for example, will recommend avoiding string interpolation because it will cause a string to be realized that may not be needed, depending on the current log-level of the application.
  4. It can never use Span or other ref struct types today, because ref structs are not allowed as generic type parameters, meaning that if a user wants to avoid copying to intermediate locations they have to manually format strings.

Internally, the runtime has a type called ValueStringBuilder to help deal with the first 2 of these scenarios. They pass a stackalloc'd buffer to the builder, repeatedly call AppendFormat with every part, and then get a final string out. If the resulting string goes past the bounds of the stack buffer, they can then move to an array on the heap. However, this type is dangerous to expose directly, as incorrect usage could lead to a rented array to be double-disposed, which then will cause all sorts of undefined behavior in the program as two locations think they have sole access to the rented array. This proposal creates a way to use this type safely from native C# code by just writing an interpolated string literal, leaving written code unchanged while improving every interpolated string that a user writes. It also extends this pattern to allow for interpolated strings passed as arguments to other methods to use a handler pattern, defined by receiver of the method, that will allow things like logging frameworks to avoid allocating strings that will never be needed, and giving C# users familiar, convenient interpolation syntax.

Detailed Design

The handler pattern

We introduce a new handler pattern that can represent an interpolated string passed as an argument to a method. The simple English of the pattern is as follows:

When an interpolated_string_expression is passed as an argument to a method, we look at the type of the parameter. If the parameter type has a constructor that can be invoked with 2 int parameters, literalLength and formattedCount, optionally takes additional parameters specified by an attribute on the original parameter, optionally has an out boolean trailing parameter, and the type of the original parameter has instance AppendLiteral and AppendFormatted methods that can be invoked for every part of the interpolated string, then we lower the interpolation using that, instead of into a traditional call to string.Format(formatStr, args). A more concrete example is helpful for picturing this:

// The handler that will actually "build" the interpolated string"
[InterpolatedStringHandler]
public ref struct TraceLoggerParamsInterpolatedStringHandler
{
    // Storage for the built-up string

    private bool _logLevelEnabled;

    public TraceLoggerParamsInterpolatedStringHandler(int literalLength, int formattedCount, Logger logger, out bool handlerIsValid)
    {
        if (!logger._logLevelEnabled)
        {
            handlerIsValid = false;
            return;
        }

        handlerIsValid = true;
        _logLevelEnabled = logger.EnabledLevel;
    }

    public void AppendLiteral(string s)
    {
        // Store and format part as required
    }

    public void AppendFormatted<T>(T t)
    {
        // Store and format part as required
    }
}

// The logger class. The user has an instance of this, accesses it via static state, or some other access
// mechanism
public class Logger
{
    // Initialization code omitted
    public LogLevel EnabledLevel;

    public void LogTrace([InterpolatedStringHandlerArguments("")]TraceLoggerParamsInterpolatedStringHandler handler)
    {
        // Impl of logging
    }
}

Logger logger = GetLogger(LogLevel.Info);

// Given the above definitions, usage looks like this:
var name = "Fred Silberberg";
logger.LogTrace($"{name} will never be printed because info is < trace!");

// This is converted to:
var name = "Fred Silberberg";
var receiverTemp = logger;
var handler = new TraceLoggerParamsInterpolatedStringHandler(literalLength: 47, formattedCount: 1, receiverTemp, out var handlerIsValid);
if (handlerIsValid)
{
    handler.AppendFormatted(name);
    handler.AppendLiteral(" will never be printed because info is < trace!");
}
receiverTemp.LogTrace(handler);

Here, because TraceLoggerParamsInterpolatedStringHandler has a constructor with the correct parameters, we say that the interpolated string has an implicit handler conversion to that parameter, and it lowers to the pattern shown above. The specese needed for this is a bit complicated, and is expanded below.

The rest of this proposal will use Append... to refer to either of AppendLiteral or AppendFormatted in cases when both are applicable.

New attributes

The compiler recognizes the System.Runtime.CompilerServices.InterpolatedStringHandlerAttribute:

using System;
namespace System.Runtime.CompilerServices
{
    [AttributeUsage(AttributeTargets.Class | AttributeTargets.Struct, AllowMultiple = false, Inherited = false)]
    public sealed class InterpolatedStringHandlerAttribute : Attribute
    {
        public InterpolatedStringHandlerAttribute()
        {
        }
    }
}

This attribute is used by the compiler to determine if a type is a valid interpolated string handler type.

The compiler also recognizes the System.Runtime.CompilerServices.InterpolatedStringHandlerArgumentAttribute:

namespace System.Runtime.CompilerServices
{
    [AttributeUsage(AttributeTargets.Parameter, AllowMultiple = false, Inherited = false)]
    public sealed class InterpolatedStringHandlerArgumentAttribute : Attribute
    {
        public InterpolatedHandlerArgumentAttribute(string argument);
        public InterpolatedHandlerArgumentAttribute(params string[] arguments);

        public string[] Arguments { get; }
    }
}

This attribute is used on parameters, to inform the compiler how to lower an interpolated string handler pattern used in a parameter position.

Interpolated string handler conversion

Type T is said to be an applicable_interpolated_string_handler_type if it is attributed with System.Runtime.CompilerServices.InterpolatedStringHandlerAttribute. There exists an implicit interpolated_string_handler_conversion to T from an interpolated_string_expression, or an additive_expression composed entirely of _interpolated_string_expression_s and using only + operators.

For simplicity in the rest of this speclet, interpolated_string_expression refers to both a simple interpolated_string_expression, and to an additive_expression composed entirely of _interpolated_string_expression_s and using only + operators.

Note that this conversion always exists, regardless of whether there will be later errors when actually attempting to lower the interpolation using the handler pattern. This is done to help ensure that there are predictable and useful errors and that runtime behavior doesn't change based on the content of an interpolated string.

Applicable function member adjustments

We adjust the wording of the applicable function member algorithm as follows (a new sub-bullet is added to each section, in bold):

A function member is said to be an applicable function member with respect to an argument list A when all of the following are true:

  • Each argument in A corresponds to a parameter in the function member declaration as described in Corresponding parameters, and any parameter to which no argument corresponds is an optional parameter.
  • For each argument in A, the parameter passing mode of the argument (i.e., value, ref, or out) is identical to the parameter passing mode of the corresponding parameter, and
    • for a value parameter or a parameter array, an implicit conversion (Implicit conversions) exists from the argument to the type of the corresponding parameter, or
    • for a ref parameter whose type is a struct type, an implicit interpolated_string_handler_conversion exists from the argument to the type of the corresponding parameter, or
    • for a ref or out parameter, the type of the argument is identical to the type of the corresponding parameter. After all, a ref or out parameter is an alias for the argument passed.

For a function member that includes a parameter array, if the function member is applicable by the above rules, it is said to be applicable in its normal form. If a function member that includes a parameter array is not applicable in its normal form, the function member may instead be applicable in its expanded form:

  • The expanded form is constructed by replacing the parameter array in the function member declaration with zero or more value parameters of the element type of the parameter array such that the number of arguments in the argument list A matches the total number of parameters. If A has fewer arguments than the number of fixed parameters in the function member declaration, the expanded form of the function member cannot be constructed and is thus not applicable.
  • Otherwise, the expanded form is applicable if for each argument in A the parameter passing mode of the argument is identical to the parameter passing mode of the corresponding parameter, and
    • for a fixed value parameter or a value parameter created by the expansion, an implicit conversion (Implicit conversions) exists from the type of the argument to the type of the corresponding parameter, or
    • for a ref parameter whose type is a struct type, an implicit interpolated_string_handler_conversion exists from the argument to the type of the corresponding parameter, or
    • for a ref or out parameter, the type of the argument is identical to the type of the corresponding parameter.

Important note: this means that if there are 2 otherwise equivalent overloads, that only differ by the type of the applicable_interpolated_string_handler_type, these overloads will be considered ambiguous. Further, because we do not see through explicit casts, it is possible that there could arise an unresolvable scenario where both applicable overloads use InterpolatedStringHandlerArguments and are totally uncallable without manually performing the handler lowering pattern. We could potentially make changes to the better function member algorithm to resolve this if we so choose, but this scenario unlikely to occur and isn't a priority to address.

Better conversion from expression adjustments

We change the better conversion from expression section to the following:

Given an implicit conversion C1 that converts from an expression E to a type T1, and an implicit conversion C2 that converts from an expression E to a type T2, C1 is a better conversion than C2 if:

  1. E is a non-constant interpolated_string_expression, C1 is an implicit_string_handler_conversion, T1 is an applicable_interpolated_string_handler_type, and C2 is not an implicit_string_handler_conversion, or
  2. E does not exactly match T2 and at least one of the following holds:

This does mean that there are some potentially non-obvious overload resolution rules, depending on whether the interpolated string in question is a constant-expression or not. For example:

void Log(string s) { ... }
void Log(TraceLoggerParamsInterpolatedStringHandler p) { ... }

Log($""); // Calls Log(string s), because $"" is a constant expression
Log($"{"test"}"); // Calls Log(string s), because $"{"test"}" is a constant expression
Log($"{1}"); // Calls Log(TraceLoggerParamsInterpolatedStringHandler p), because $"{1}" is not a constant expression

This is introduced so that things that can simply be emitted as constants do so, and don't incur any overhead, while things that cannot be constant use the handler pattern.

InterpolatedStringHandler and Usage

We introduce a new type in System.Runtime.CompilerServices: DefaultInterpolatedStringHandler. This is a ref struct with many of the same semantics as ValueStringBuilder, intended for direct use by the C# compiler. This struct would look approximately like this:

// API Proposal issue: https://github.com/dotnet/runtime/issues/50601
namespace System.Runtime.CompilerServices
{
    [InterpolatedStringHandler]
    public ref struct DefaultInterpolatedStringHandler
    {
        public DefaultInterpolatedStringHandler(int literalLength, int formattedCount);
        public string ToStringAndClear();

        public void AppendLiteral(string value);

        public void AppendFormatted<T>(T value);
        public void AppendFormatted<T>(T value, string? format);
        public void AppendFormatted<T>(T value, int alignment);
        public void AppendFormatted<T>(T value, int alignment, string? format);

        public void AppendFormatted(ReadOnlySpan<char> value);
        public void AppendFormatted(ReadOnlySpan<char> value, int alignment = 0, string? format = null);

        public void AppendFormatted(string? value);
        public void AppendFormatted(string? value, int alignment = 0, string? format = null);

        public void AppendFormatted(object? value, int alignment = 0, string? format = null);
    }
}

We make a slight change to the rules for the meaning of an interpolated_string_expression:

If the type of an interpolated string is string and the type System.Runtime.CompilerServices.DefaultInterpolatedStringHandler exists, and the current context supports using that type, the string is lowered using the handler pattern. The final string value is then obtained by calling ToStringAndClear() on the handler type. Otherwise, if the type of an interpolated string is System.IFormattable or System.FormattableString [the rest is unchanged]

The "and the current context supports using that type" rule is intentionally vague to give the compiler leeway in optimizing usage of this pattern. The handler type is likely to be a ref struct type, and ref struct types are normally not permitted in async methods. For this particular case, the compiler would be allowed to make use the handler if none of the interpolation holes contain an await expression, as we can statically determine that the handler type is safely used without additional complicated analysis because the handler will be dropped after the interpolated string expression is evaluated.

Open Question:

Do we want to instead just make the compiler know about DefaultInterpolatedStringHandler and skip the string.Format call entirely? It would allow us to hide a method that we don't necessarily want to put in people's faces when they manually call string.Format.

Answer: Yes.

Open Question:

Do we want to have handlers for System.IFormattable and System.FormattableString as well?

Answer: No.

Handler pattern codegen

In this section, method invocation resolution refers to the steps listed here.

Constructor resolution

Given an applicable_interpolated_string_handler_type T and an interpolated_string_expression i, method invocation resolution and validation for a valid constructor on T is performed as follows:

  1. Member lookup for instance constructors is performed on T. The resulting method group is called M.
  2. The argument list A is constructed as follows:
    1. The first two arguments are integer constants, representing the literal length of i, and the number of interpolation components in i, respectively.
    2. If i is used as an argument to some parameter pi in method M1, and parameter pi is attributed with System.Runtime.CompilerServices.InterpolatedStringHandlerArgumentAttribute, then for every name Argx in the Arguments array of that attribute the compiler matches it to a parameter px that has the same name. The empty string is matched to the receiver of M1.
      • If any Argx is not able to be matched to a parameter of M1, or an Argx requests the receiver of M1 and M1 is a static method, an error is produced and no further steps are taken.
      • Otherwise, the type of every resolved px is added to the argument list, in the order specified by the Arguments array. Each px is passed with the same ref semantics as is specified in M1.
    3. The final argument is a bool, passed as an out parameter.
  3. Traditional method invocation resolution is performed with method group M and argument list A. For the purposes of method invocation final validation, the context of M is treated as a member_access through type T.
    • If a single-best constructor F was found, the result of overload resolution is F.
    • If no applicable constructors were found, step 3 is retried, removing the final bool parameter from A. If this retry also finds no applicable members, an error is produced and no further steps are taken.
    • If no single-best method was found, the result of overload resolution is ambiguous, an error is produced, and no further steps are taken.
  4. Final validation on F is performed.
    • If any element of A occurred lexically after i, an error is produced and no further steps are taken.
    • If any A requests the receiver of F, and F is an indexer being used as an initializer_target in a member_initializer, then an error is reported and no further steps are taken.

Note: the resolution here intentionally do not use the actual expressions passed as other arguments for Argx elements. We only consider the types post-conversion. This makes sure that we don't have double-conversion issues, or unexpected cases where a lambda is bound to one delegate type when passed to M1 and bound to a different delegate type when passed to M.

Note: We report an error for indexers uses as member initializers because of the order of evaluation for nested member initializers. Consider this code snippet:


var x1 = new C1 { C2 = { [GetString()] = { A = 2, B = 4 } } };

/* Lowering:
__c1 = new C1();
string argTemp = GetString();
__c1.C2[argTemp][1] = 2;
__c1.C2[argTemp][3] = 4;

Prints:
GetString
get_C2
get_C2
*/

string GetString()
{
    Console.WriteLine("GetString");
    return "";
}

class C1
{
    private C2 c2 = new C2();
    public C2 C2 { get { Console.WriteLine("get_C2"); return c2; } set { } }
}

class C2
{
    public C3 this[string s]
    {
        get => new C3();
        set { }
    }
}

class C3
{
    public int A
    {
        get => 0;
        set { }
    }
    public int B
    {
        get => 0;
        set { }
    }
}

The arguments to __c1.C2[] are evaluated before the receiver of the indexer. While we could come up with a lowering that works for this scenario (either by creating a temp for __c1.C2 and sharing it across both indexer invocations, or only using it for the first indexer invocation and sharing the argument across both invocations) we think that any lowering would be confusing for what we believe is a pathological scenario. Therefore, we forbid the scenario entirely.

Open Question:

If we use a constructor instead of Create, we'd improve runtime codegen, at the expense of narrowing the pattern a bit.

Answer: We will restrict to constructors for now. We can revisit adding a general Create method later if the scenario arises.

Append... method overload resolution

Given an applicable_interpolated_string_handler_type T and an interpolated_string_expression i, overload resolution for a set of valid Append... methods on T is performed as follows:

  1. If there are any interpolated_regular_string_character components in i:
    1. Member lookup on T with the name AppendLiteral is performed. The resulting method group is called Ml.
    2. The argument list Al is constructed with one value parameter of type string.
    3. Traditional method invocation resolution is performed with method group Ml and argument list Al. For the purposes of method invocation final validation, the context of Ml is treated as a member_access through an instance of T.
      • If a single-best method Fi is found and no errors were produced, the result of method invocation resolution is Fi.
      • Otherwise, an error is reported.
  2. For every interpolation ix component of i:
    1. Member lookup on T with the name AppendFormatted is performed. The resulting method group is called Mf.
    2. The argument list Af is constructed:
      1. The first parameter is the expression of ix, passed by value.
      2. If ix directly contains a constant_expression component, then an integer value parameter is added, with the name alignment specified.
      3. If ix is directly followed by an interpolation_format, then a string value parameter is added, with the name format specified.
    3. Traditional method invocation resolution is performed with method group Mf and argument list Af. For the purposes of method invocation final validation, the context of Mf is treated as a member_access through an instance of T.
      • If a single-best method Fi is found, the result of method invocation resolution is Fi.
      • Otherwise, an error is reported.
  3. Finally, for every Fi discovered in steps 1 and 2, final validation is performed:
    • If any Fi does not return bool by value or void, an error is reported.
    • If all Fi do not return the same type, an error is reported.

Note that these rules do not permit extension methods for the Append... calls. We could consider enabling that if we choose, but this is analogous to the enumerator pattern, where we allow GetEnumerator to be an extension method, but not Current or MoveNext().

These rules do permit default parameters for the Append... calls, which will work with things like CallerLineNumber or CallerArgumentExpression (when supported by the language).

We have separate overload lookup rules for base elements vs interpolation holes because some handlers will want to be able to understand the difference between the components that were interpolated and the components that were part of the base string.

Open Question

Some scenarios, like structured logging, want to be able to provide names for interpolation elements. For example, today a logging call might look like Log("{name} bought {itemCount} items", name, items.Count);. The names inside the {} provide important structure information for loggers that help with ensuring output is consistent and uniform. Some cases might be able to reuse the :format component of an interpolation hole for this, but many loggers already understand format specifiers and have existing behavior for output formatting based on this info. Is there some syntax we can use to enable putting these named specifiers in?

Some cases may be able to get away with CallerArgumentExpression, provided that support does land in C# 10. But for cases that invoke a method/property, that may not be sufficient.

Answer:

While there are some interesting parts to templated strings we could explore in an orthogonal language feature, we don't think a specific syntax here has much benefit over solutions such as using a tuple: $"{("StructuredCategory", myExpression)}".

Performing the conversion

Given an applicable_interpolated_string_handler_type T and an interpolated_string_expression i that had a valid constructor Fc and Append... methods Fa resolved, lowering for i is performed as follows:

  1. Any arguments to Fc that occur lexically before i are evaluated and stored into temporary variables in lexical order. In order to preserve lexical ordering, if i occurred as part of a larger expression e, any components of e that occurred before i will be evaluated as well, again in lexical order.
  2. Fc is called with the length of the interpolated string literal components, the number of interpolation holes, any previously evaluated arguments, and a bool out argument (if Fc was resolved with one as the last parameter). The result is stored into a temporary value ib.
    1. The length of the literal components is calculated after replacing any open_brace_escape_sequence with a single {, and any close_brace_escape_sequence with a single }.
  3. If Fc ended with a bool out argument, a check on that bool value is generated. If true, the methods in Fa will be called. Otherwise, they will not be called.
  4. For every Fax in Fa, Fax is called on ib with either the current literal component or interpolation expression, as appropriate. If Fax returns a bool, the result is logically anded with all preceeding Fax calls.
    1. If Fax is a call to AppendLiteral, the literal component is unescaped by replacing any open_brace_escape_sequence with a single {, and any close_brace_escape_sequence with a single }.
  5. The result of the conversion is ib.

Again, note that arguments passed to Fc and arguments passed to e are the same temp. Conversions may occur on top of the temp to convert to a form that Fc requires, but for example lambdas cannot be bound to a different delegate type between Fc and e.

Open Question

This lowering means that subsequent parts of the interpolated string after a false-returning Append... call don't get evaluated. This could potentially be very confusing, particularly if the format hole is side-effecting. We could instead evaluate all format holes first, then repeatedly call Append... with the results, stopping if it returns false. This would ensure that all expressions get evaluated as one might expect, but we call as few methods as we need to. While the partial evaluation might be desirable for some more advanced cases, it is perhaps non-intuitive for the general case.

Another alternative, if we want to always evaluate all format holes, is to remove the Append... version of the API and just do repeated Format calls. The handler can track whether it should just be dropping the argument and immediately returning for this version.

Answer: We will have conditional evaluation of the holes.

Open Question

Do we need to dispose of disposable handler types, and wrap calls with try/finally to ensure that Dispose is called? For example, the interpolated string handler in the bcl might have a rented array inside it, and if one of the interpolation holes throws an exception during evaluation, that rented array could be leaked if it wasn't disposed.

Answer: No. handlers can be assigned to locals (such as MyHandler handler = $"{MyCode()};), and the lifetime of such handlers is unclear. Unlike foreach enumerators, where the lifetime is obvious and no user-defined local is created for the enumerator.

Other considerations

Allow string types to be convertible to handlers as well

For type author simplicity, we could consider allowing expressions of type string to be implicitly-convertible to applicable_interpolated_string_handler_types. As proposed today, authors will likely need to overload on both that handler type and regular string types, so their users don't have to understand the difference. This may be an annoying and non-obvious overhead, as a string expression can be viewed as an interpolation with expression.Length prefilled length and 0 holes to be filled.

This would allow new APIs to only expose a handler, without also having to expose a string-accepting overload. However, it won't get around the need for changes to better conversion from expression, so while it would work it may be unnecessary overhead.

Answer:

We think that this could end up being confusing, and there's an easy workaround for custom handler types: add a user-defined conversion from string.

Incorporating spans for heap-less strings

ValueStringBuilder as it exists today has 2 constructors: one that takes a count, and allocates on the heap eagerly, and one that takes a Span<char>. That Span<char> is usually a fixed size in the runtime codebase, around 250 elements on average. To truly replace that type, we should consider an extension to this where we also recognize GetInterpolatedString methods that take a Span<char>, instead of just the count version. However, we see a few potential thorny cases to resolve here:

  • We don't want to stackalloc repeatedly in a hot loop. If we were to do this extension to the feature, we'd likely want to share the stackalloc'd span between loop iterations. We know this is safe, as Span<T> is a ref struct that can't be stored on the heap, and users would have to be pretty devious to manage to extract a reference to that Span (such as creating a method that accepts such a handler then deliberately retrieving the Span from the handler and returning it to the caller). However, allocating ahead of time produces other questions:
    • Should we eagerly stackalloc? What if the loop is never entered, or exits before it needs the space?
    • If we don't eagerly stackalloc, does that mean we introduce a hidden branch on every loop? Most loops likely won't care about this, but it could affect some tight loops that don't want to pay the cost.
  • Some strings can be quite big, and the appropriate amount to stackalloc is dependent on a number of factors, including runtime factors. We don't really want the C# compiler and specification to have to determine this ahead of time, so we'd want to resolve https://github.com/dotnet/runtime/issues/25423 and add an API for the compiler to call in these cases. It also adds more pros and cons to the points from the previous loop, where we don't want to potentially allocate large arrays on the heap many times or before one is needed.

Answer:

This is out of scope for C# 10. We can look at this in general when we look at the more general params Span<T> feature.

Non-try version of the API

For simplicity, this spec currently just proposes recognizing a Append... method, and things that always succeed (like InterpolatedStringHandler) would always return true from the method. This was done to support partial formatting scenarios where the user wants to stop formatting if an error occurs or if it's unnecessary, such as the logging case, but could potentially introduce a bunch of unnecessary branches in standard interpolated string usage. We could consider an addendum where we use just FormatX methods if no Append... method is present, but it does present questions about what we do if there's a mix of both Append... and FormatX calls.

Answer:

We want the non-try version of the API. The proposal has been updated to reflect this.

Passing previous arguments to the handler

There is unfortunate lack of symmetry in the proposal at it currently exists: invoking an extension method in reduced form produces different semantics than invoking the extension method in normal form. This is different from most other locations in the language, where reduced form is just a sugar. We propose adding an attribute to the framework that we will recognize when binding a method, that informs the compiler that certain parameters should be passed to the constructor on the handler. Usage looks like this:

namespace System.Runtime.CompilerServices
{
    [AttributeUsage(AttributeTargets.Parameter, AllowMultiple = false, Inherited = false)]
    public sealed class InterpolatedStringHandlerArgumentAttribute : Attribute
    {
        public InterpolatedStringHandlerArgumentAttribute(string argument);
        public InterpolatedStringHandlerArgumentAttribute(params string[] arguments);

        public string[] Arguments { get; }
    }
}

Usage of this is then:

namespace System
{
    public sealed class String
    {
        public static string Format(IFormatProvider? provider, [InterpolatedStringHandlerArgument("provider")] ref DefaultInterpolatedStringHandler handler);
        …
    }
}

namespace System.Runtime.CompilerServices
{
    public ref struct DefaultInterpolatedStringHandler
    {
        public DefaultInterpolatedStringHandler(int baseLength, int holeCount, IFormatProvider? provider); // additional factory
        …
    }
}

var formatted = string.Format(CultureInfo.InvariantCulture, $"{X} = {Y}");

// Is lowered to

var tmp1 = CultureInfo.InvariantCulture;
var handler = new DefaultInterpolatedStringHandler(3, 2, tmp1);
handler.AppendFormatted(X);
handler.AppendLiteral(" = ");
handler.AppendFormatted(Y);
var formatted = string.Format(tmp1, handler);

The questions we need to answer:

  1. Do we like this pattern in general?
  2. Do we want to allow these arguments to come from after the handler parameter? Some existing patterns in the BCL, such as Utf8Formatter, put the value to be formatted before the thing needed to format into. To fit in best with these patterns, we'd likely want to allow this, but we need to decide if this out-of-order evaluate is ok.

Answer:

We want to support this. The spec has been updated to reflect this. Arguments will be required to be specified in lexical order at the call site, and if a needed argument to the create method is specified after the interpolated string literal, an error is produced.

await usage in interpolation holes

Because $"{await A()}" is a valid expression today, we need to rationalize how interpolation holes with await. We could solve this with a few rules:

  1. If an interpolated string used as a string, IFormattable, or FormattableString has an await in an interpolation hole, fall back to old-style formatter.
  2. If an interpolated string is subject to an implicit_string_handler_conversion and applicable_interpolated_string_handler_type is a ref struct, await is not allowed to be used in the format holes.

Fundamentally, this desugaring could use a ref struct in an async method as long as we guarantee that the ref struct will not need to be saved to the heap, which should be possible if we forbid awaits in the interpolation holes.

Alternatively, we could simply make all handler types non-ref structs, including the framework handler for interpolated strings. This would, however, preclude us from someday recognizing a Span version that does not need to allocate any scratch space at all.

Answer:

We will treat interpolated string handlers the same as any other type: this means that if the handler type is a ref struct and the current context doesn't allow the usage of ref structs, it is illegal to use handler here. The spec around lowering of string literals used as strings is intentionally vague to allow the compiler to decide on what rules it deems appropriate, but for custom handler types they will have to follow the same rules as the rest of the language.

Handlers as ref parameters

Some handlers might want to be passed as ref parameters (either in or ref). Should we allow either? And if so, what will a ref handler look like? ref $"" is confusing, as you're not actually passing the string by ref, you're passing the handler that is created from the ref by ref, and has similar potential issues with async methods.

Answer:

We want to support this. The spec has been updated to reflect this. The rules should reflect the same rules that apply to extension methods on value types.

Interpolated strings through binary expressions and conversions

Because this proposal makes interpolated strings context sensitive, we would like to allow the compiler to treat a binary expression composed entirely of interpolated strings, or an interpolated string subjected to a cast, as an interpolated string literal for the purposes of overload resolution. For example, take the following scenario:

struct Handler1
{
    public Handler1(int literalLength, int formattedCount, C c) => ...;
    // AppendX... methods as necessary
}
struct Handler2
{
    public Handler2(int literalLength, int formattedCount, C c) => ...;
    // AppendX... methods as necessary
}

class C
{
    void M(Handler1 handler) => ...;
    void M(Handler2 handler) => ...;
}

c.M($"{X}"); // Ambiguous between the M overloads

This would be ambiguous, necessitating a cast to either Handler1 or Handler2 in order to resolve. However, in making that cast, we would potentially throw away the information that there is context from the method receiver, meaning that the cast would fail because there is nothing to fill in the information of c. A similar issue arises with binary concatenation of strings: the user could want to format the literal across several lines to avoid line wrapping, but would not be able to because that would no longer be an interpolated string literal convertible to the handler type.

To resolve these cases, we make the following changes:

  • An additive_expression composed entirely of interpolated_string_expressions and using only + operators is considered to be an interpolated_string_literal for the purposes of conversions and overload resolution. The final interpolated string is created by logically concatinating all individual interpolated_string_expression components, from left to right.
  • A cast_expression or a relational_expression with operator as whose operand is an interpolated_string_expressions is considered an interpolated_string_expressions for the purposes of conversions and overload resolution.

Open Questions:

Do we want to do this? We don't do this for System.FormattableString, for example, but that can be broken out onto a different line, whereas this can be context-dependent and therefore not able to be broken out into a different line. There are also no overload resolution concerns with FormattableString and IFormattable.

Answer:

We think that this is a valid use case for additive expressions, but that the cast version is not compelling enough at this time. We can add it later if necessary. The spec has been updated to reflect this decision.

Other use cases

See https://github.com/dotnet/runtime/issues/50635 for examples of proposed handler APIs using this pattern.