Skip to content

Latest commit

 

History

History
294 lines (211 loc) · 13.5 KB

expand-ref.md

File metadata and controls

294 lines (211 loc) · 13.5 KB

Expanding ref support

Summary

This proposal expands the capabilities of ref and scoped in the language. The goal being to leverage the existing types of rules in the model to allow ref struct usage in more locations and provide more lifetime expressiveness for APIs.

Motivation

There are still a number of scenarios around ref which cannot be safely expressed in the language. These are generally when using multiple mutable ref struct parameters where many are passed by ref or when trying to use ref struct in ref fields.

To fully satisfy all of these scenarios would require us to introduce explicit lifetime parameters and relationships into the language. That is a huge investment that is not yet motivated by need. Instead this proposal takes our existing lifetime annotation, scoped, and sees how much further ref safety can be taken without introducing any other annotations or keywords.

This doesn't solve all scenarios but does remove several known friction points in the language. It also serves to show us exactly where the limits are without introducing explicit lifetime parameters.

Detailed Design

The rules for ref struct safety are defined in the following documents:

This proposal will be building on top of those previous ones.

The more detailed rules will rely on the annotation syntax to describe the detailed rules. This is the most direct way to discuss how syntax behaves in the greater model. Readers interested in the very low level details should familiarize themselves with that syntax before digesting this proposal.

ref scoped parameters

The language will allow for parameters to be declared as ref scoped. This will serve to constrain the safe-to-escape of the value such that it cannot be returned from the current method.

Span<int>M(Span<T>p1,ref scoped Span<int>p2){// Error: cannot return scoped valuereturnp2;// Error: the safe-to-escape of p1 is not convertible to p2.p2=p1;// Okay: heap can always be assignedp2=default;// Okayp2[0]=42;}

This capability will help cases where multiple ref struct values with different lifetimes are passed by ref. Having ref scoped allows developers to note which values do not escape and that allows for more call site flexibility.

refstructData{ ...}voidCopy1(refDatasource,refDatadest){ ...}voidCopy2(refDatasource,ref scoped Datadest){ ...}voidUse(refDatadata){// STE: current methodvarlocal=newData(stackallocint[42]);// Error: compiler has to assume local copied to data Copy1(refdata,reflocal);// Okay: compiler knows lifetime only flows data -> localCopy2(refdata,reflocal);}

This is accomplished by giving every ref scoped parameter a new escape scope named current parameter N where N is the numeric order of the parameter. For example the first parameter has a safe-to-escape of current parameter 1. An escape scope of current parameter N can be converted to current method but has no other defined relationship. That serves to restrict their usage to the current method.

It's important to note each parameter has a different current parameter N scope. That means they cannot be assigned to each other. This is necessary to prevent ref scoped parameters from returning each others data.

voidSwap(ref scoped Span<int>p1,ref scoped Span<int>p2){// Error: can't assign current parameter 2 to current parameter 1p2=p1;// Error: can't assign current parameter 1 to current parameter 2p1=p2;// Okay: as current parameter 1 and 2 can be converted to current method scoped Span<int>local1=p1; scoped Span<int>local2=p2;// Okay: however the safe-to-escape here is current parameter N, not // current method so this could cause a bit of confusion later onSpan<int>local3=p1;Span<int>local4=p2;// Okay: the safe-to-escape of the value is inferred in this case as it is // done for ref locals today.refSpan<int>refLocal1=refp1;refSpan<int>refLocal2=refp2;}

A ref scoped parameter is also implicitly scoped ref. That means neither the value nor its ref can be returned from the method. Both ref and in parameters can have their values modified with scoped. An out parameter cannot have its value modified with scoped as such a declaration is non-sensical.

voidM(ref scoped Span<int>p1,// Okayin scoped Span<int>p2,// Okayout scoped Span<int>p2,// Error)

The method arguments must match rules will be updated to take ref scoped into account. Values passed to such parameters do not need to be considered when calculating the return scopes.

Detailed notes:

  • A ref scoped parameter is implicitly scoped ref
  • An out scoped parameter declaration is an error

ref field to ref struct

The language will allow for ref struct to appear as ref scoped fields. This scoped will serve to ensure the values cannot be escaped outside the containing instance but can be read and manipulated within it.

refstructDeserializer{ref scoped Utf8JsonReaderreader;ReadOnlySpan<byte>M1(){// okay: implicitly scoped to current methodvarspan=reader.ValueSpan;// okayreader.Skip();// Error: can't escape the ref data the ref scoped field refers toreturnreader.ValueSpan;}}

This is accomplished by giving every ref scoped field two new escape scopes named current field N and current ref field N where N is the numeric order of the field. For example, the first field has a safe-to-escape of current field 1 and a ref-safe-to-escape of current ref field N. Both escape scopes can be converted to current method, and current field N can be converted to current ref field N, but no other defined relationships exist. That serves to restrict their usage to the current method where the containing value is used. This escape scope applies to both.

Below are a few examples of these rules in action

refstructNestedRefStruct{}refstructRefStruct{publicNestedRefStructNestedField;}refstructS{ref scoped RefStructfield;RefStructM1(RefStructs){// Okayfield=new();// Error: calling-method is not convertible to current-field-1 as they have // no relationshipfield=s;// Error: safe-to-escape is current-field-1 which isn't returnable returnfield;}NestedRefStructM2(){// Error: safe-to-escape is current-field-1 which isn't returnable returnfield.NestedField;}refRefStructM3(){// Error: safe-to-escape is current-ref-field-1 which isn't returnable returnreffield;}}

The method arguments must match rules do not need to be updated here as they already account for ref parameters being captured as ref field. Even though a ref to ref struct was not directly returnable before, it could be returned indirectly by a ref to a struct field of the value.

The language will also allow for ref fields to be declared as scoped ref. There are less use cases for this but ref scoped implies scoped ref hence the rules must be adjusted to account for this. As such the syntax will be exposed because while the use cases are small the infrastructure already exists. The ref-safe-to-escape of such fields follows the logic above for ref scoped fields.

Detailed notes:

  • A ref field where the type is a ref struct must be ref scoped
  • A ref field may be marked scoped ref

Sunset restricted types

The ability for any type to be a ref field allows us to fully sunset the notion of restricted types. The compiler has a concept of a set of restricted types which is largely undocumented. These types were given a special status because in C# 1.0 there was no general purpose way to express their behavior. Most notably the fact that the types can contain references to the execution stack. Instead the compiler had special knowledge of them and restricted their use to ways that would always be safe: disallowed returns, cannot use as array elements, cannot use in generics, etc ...

Once ref fields are available and extended to support ref struct these types can be fully rationalized within those rules. As such the compiler will no longer have the notion of restricted types when using a language version that supports ref fields of ref struct.

To support this our ref safety rules will be updated as follows:

  • __makeref(e) will be logically treated as a method with the signature static TypedReference __makeref(ref T value) were T is the type of e.
  • __refvalue(e, T)
    • When T is a ref struct: will be treated as accessing a field declared as ref scoped T inside e.
    • Will be treated as accessing a field declared as ref T inside e
  • __arglist as a parameter will be implicitly scoped
  • __arglist(...) as an expression will have a ref-safe-to-escape and safe-to-escape of current method.

Conforming runtimes will ensure that TypedReference, RuntimeArgumentHandle and ArgIterator are defined as ref struct. Further TypedReference must be viewed as having a ref field to a ref struct for any possible type (it can store any value). That combined with the above rules will ensure references to the stack do not escape beyond their lifetime.

Note: strictly speaking this is a compiler implementation detail vs. part of the language. But given the relationship with ref fields it is being included in the language proposal for simplicity.

Annotation Definition

At an annotation level every parameter marked ref scoped will have a new lifetime parameter defined. The name will be $paramN where N is the numerical order of the parameter. That lifetime will only have the relationship where $paramN : $local.

refstructS{}voidM(ref scoped Ss)// maps to  void M<$param1>(ref<$local> S<$param1> s)where $param1: $local

This definition prevents the value from escaping from the method as the lifetime is not returnable. It also prevents local data from escaping from the current method through the parameter as the lifetime is wider than $local but not equivalent.

voidM<$param1>(ref<$local>S<$param1>p)where $param1: $local{S<$local>s=newS<$local>(stackallocint[42]);// error: cannot convert S<$local> to S<$param1>p=s;}

At an annotation level every field marked scoped ref (explicitly or implicitly via ref scoped) will have a new lifetime parameter defined. The name will be $refFieldN where N is the numerical order of the field. That lifetime will have the relationship where $refFieldN : $local in all methods that use the type.

refstructS{ scoped refinti;}SM(Sp){}// maps to refstructS<out $this, $refField1>{ref<$refField1>inti;}S<$cm>M<$cm, $l1>(S<$cm, $l1>p)where $l1: $local{}

Every field marked as ref scoped will have a new lifetime parameter defined. The name will be $fieldN where N is the numerical order of the field. That lifetime will have the relationship where $fieldN : $refFieldN defined on the type. It will also have the relationship where $fieldN : $local in all method that use the type.

refstructS1{}refstructS2{ref scoped S1field;}S2M(S2p){}// maps to refstructS1<out $this>{}refstructS2<out $this, $refField1, $field1>where $field1: $refField1{ref<$refField1>S1<$field1>field;}S1<$cm, $l1, $l2>M<$cm, $l1>M(S<$cm, $l1, $l2>p)where $l2: $l1where $l1: $local{}

These definitions prevent the values (ref or value) from escaping as their lifetimes are never returnable. It does allow for them to be manipulated and adjusted though. Non ref data, or data known to have $heap lifetime, can be assigned into such fields.

Open Issues

Ability to mark this as ref scoped

The proposal does not provide any way to mark this as ref scoped for a given method. At this time the author can see no significant benefits to this. If such scenarios do come along then an attribute such as [RefScoped] could be introduced similar to how [UnscopedRef] works.

Requiring ref fields to ref struct to be scoped

Certain readers are likely to be disappointed that ref field to ref struct must be ref scoped. That limits the number of scenarios which can assign ref data into such fields.

This is unfortunately necessary given the constraints of the design. Having a plain ref effectively requires that explicit lifetime annotations exist in the language. There is no other way to safely express the relationship between the value and the container.

close