Skip to content

Commit

Permalink
Update docs for ByRefLike with generics for work in .NET 10 (#103318)
Browse files Browse the repository at this point in the history
Co-authored-by: Jan Kotas <jkotas@microsoft.com>
  • Loading branch information
2 people authored and directhex committed Jul 26, 2024
1 parent 4541933 commit 4ff0041
Show file tree
Hide file tree
Showing 6 changed files with 447 additions and 113 deletions.
138 changes: 121 additions & 17 deletions docs/design/features/byreflike-generics.md
Original file line number Diff line number Diff line change
Expand Up @@ -7,22 +7,26 @@ Using ByRefLike types in Generic parameters is possible by building upon support

## Runtime impact

Supporting ByRefLike type as Generic parameters will impact the following IL instructions:
Supporting ByRefLike types as Generic parameters will impact the following IL instructions.

- `box` &ndash; Types with ByRefLike parameters used in fields cannot be boxed.
The `constrained. callvirt` sequence is valid if a ByRefLike type is provided. A `NotSupportedException` will be thrown at the call-site, if the target resolves to a method implemented on `object` or a default interface method.

Throws `InvalidProgramException` when passed a ByRefLike type:
- `box` &ndash; ByRefLike types cannot be allocated on the heap.

Throws `TypeLoadException` when passed a ByRefLike type:
- `stsfld` / `ldsfld` &ndash; Type fields of a ByRefLike parameter cannot be marked `static`.
- `newarr` / `stelem` / `ldelem` / `ldelema` &ndash; Arrays are not able to contain ByRefLike types.
- `newobj` &ndash; For multi-dimensional array construction.
- `constrained.callvirt` &ndash; If this IL sequence resolves to a method implemented on `object` or default interface method, an error will occur during the attempt to box the instance.

If any of the above instructions are attempted to be used with a ByRefLike type, the runtime will throw an `InvalidProgramException`. Sequences involving some of the above instructions are considered optimizations and represent cases that will remain valid regardless of a `T` being ByRefLike. See "Special IL Sequences" section below for details.

The following instructions are already set up to support this feature since their behavior will fail as currently defined due to the inability to box a ByRefLike type.

- `throw` &ndash; Requires an object reference to be on stack, which can never be a ByRefLike type.
- `unbox` / `unbox.any` &ndash; Requires an object reference to be on stack, which can never be a ByRefLike type.
- `isinst` &ndash; Will always place `null` on stack.
- `castclass` &ndash; Will always throw `InvalidCastException`.
- `throw`
- `unbox` / `unbox.any`
- `isinst`
- `castclass`

**NOTE** There are sequences involving some of the above instructions that may remain valid regardless of a `T` being ByRefLike&mdash;see ["Options for invalid IL" section](#invalid_il_options) below for details.

The expansion of ByRefLike types as Generic parameters does not relax restrictions on where ByRefLike types can be used. When `T` is ByRefLike, the use of `T` as a field will require the enclosing type to be ByRefLike.

Expand Down Expand Up @@ -110,23 +114,123 @@ throw

Adding `gpAcceptByRefLike` to the metadata of a Generic parameter will be considered a non-breaking binary change.

Enumerating of constructors/methods on `Span<T>` and `ReadOnlySpan<T>` may throw `TypeLoadException` if `T` is a ByRefLike type. See "Troublesome APIs" above for the list of APIs that cause this condition.
Enumerating of constructors/methods on `Span<T>` and `ReadOnlySpan<T>` may throw `TypeLoadException` if `T` is a ByRefLike type. See "Troublesome API mitigation" above for the list of APIs that cause this condition.

## <a name="invalid_il_options"></a> Options for invalid IL

There are two potential options below for how to address this issue. Based on communication with the Roslyn team, option (1) is the current plan of record for .NET 10.

The first indented IL sequences below represents the `is-type` sequence. Combining the first with the second indented section represents the "type pattern matching" scenario in C#. The below sequence performs a type check and then, if successful, consumes the unboxed instance.

```IL
// Type check
ldarg.0
box <Source>
isinst <Target>
brfalse.s NOT_INST
// Unbox and store unboxed instance
ldarg.0
box <Source>
isinst <Target>
unbox.any <Target>
stloc.X
NOT_INST:
ret
```

With the above IL composition implemented, the following C# describes the following "type pattern matching" scenarios and what one might expect given current C# semantics.

```csharp
struct S {}
struct S<T> {}
ref struct RS {}
ref struct RS<T> {}
interface I {}
class C {}
class C<T> {}

// Not currently valid C#
void M<T, U>(T t) where T: allows ref struct
{
// Valid
if (t is int i)

if (t is S s)
if (t is S<char> sc)
if (t is S<U> su)

if (t is RS rs)
if (t is RS<char> rsc)
if (t is RS<U> rsu)

if (t is string str)
if (t is C c)
if (t is C<I> ci)
if (t is C<U> cu)

// Can be made to work in IL.
if (t is I itf) // A new local "I" would not be used for ByRefLike scenarios.
// The local would be the ByRefLike type, not "I".
// Invalid
if (t is object o) // ByRefLike types evaluate "true" for object.
if (t is U u)
}
```

### Option 1) Compiler helpers

The following two helper functions could be introduced and would replace currently invalid `is-type` IL sequences when ByRefLike types are involved. Their behavior would broadly be defined to operate as if the ByRefLike aspect of either the `TFrom` and `TTo` is not present. An alternative approach would be consult with the Roslyn team and define the semantics of these functions to adhere to C# language rules.

```csharp
namespace System.Runtime.CompilerServices
{
public static class RuntimeHelpers
{
// Replacement for the [box; isinst; brfalse/true] sequence.
public static bool IsInstanceOf<TFrom, TTo>(TFrom source)
where TFrom: allows ref struct
where TTo: allows ref struct;

// Replacement for the [box; isinst; unbox.any] sequence.
// Would throw InvalidCastException for invalid use at run-time.
// For example:
// TFrom: RS, TTo: object => always throws
// TFrom: RS, TTo: <interface> => always throws
public static TTo CastTo<TFrom, TTo>(TFrom source)
where TFrom: allows ref struct
where TTo: allows ref struct;
}
}
```

Example usage of the above methods.

```csharp
TTo result;
if (RuntimeHelpers.IsInstanceOf<TFrom, TTo>(source))
{
result = RuntimeHelpers.CastTo<TFrom, TTo>(source);
}
```

## Special IL Sequences
### Option 2) Special IL sequences

The following are IL sequences involving the `box` instruction. They are used for common C# language constructs and shall continue to be valid, even with ByRefLike types, in cases where the result can be computed at JIT time and elided safely. These sequences must now be elided when the target type is ByRefLike. The conditions where each sequence is elided are described below and each condition will be added to the ECMA-335 addendum.
The following are IL sequences involving the `box` instruction. They are used for common C# language constructs and would continue to be valid, even with ByRefLike types. These sequences would be **required** to be valid when the target type is ByRefLike. Each sequence would be added to the ECMA-335 addendum.

`box` ; `unbox.any` &ndash; The box target type is equal to the unboxed target type.
`box` ; `isinst` ; `br_true/false` &ndash; Passing a ByRefLike type as the argument to the `box` instruction is permitted to accomplish a type check, in C# `x is Y`. **Note** ByRefLike types would evaluate to `true` when compared against `System.Object`.

`box` ; `br_true/false` &ndash; The box target type is non-`Nullable<T>`.
`box` ; `isinst` ; `unbox.any` &ndash; In order to permit "type pattern matching", in C# `x is Y y`, this sequence will permit use of a ByRefLike type on any instruction, but does not permit the use of generic parameters being exposed to `isinst` or `unbox.any`.

`box` ; `isinst` ; `unbox.any` &ndash; The box, `isint`, and unbox target types are all equal.
`box` ; `unbox.any` &ndash; Valid to use ByRefLike types.

`box` ; `isinst` ; `br_true/false` &ndash; The box target type is equal to the unboxed target type or the box target type is `Nullable<T>` and target type equalities can be computed.
`box` ; `br_true/false` &ndash; Valid to use ByRefLike types.

## Examples

Below are valid and invalid examples of ByRefLike as Generic parameters. All examples use the **not official** syntax, `allows ref struct`, for indicating the Generic permits ByRefLike types.
Below are currently (.NET 9) valid and invalid examples of ByRefLike as Generic parameters.

**1) Valid**
```csharp
Expand Down
28 changes: 27 additions & 1 deletion docs/design/specs/Ecma-335-Augments.md
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,8 @@ This is a list of additions and edits to be made in ECMA-335 specifications. It
- [Covariant Return Types](#covariant-return-types)
- [Function Pointer Type Identity](#function-pointer-type-identity)
- [Unsigned data conversion with overflow detection](#unsigned-data-conversion-with-overflow-detection)
- [Ref field support](#ref-fields)
- [Ref fields support](#ref-fields)
- [ByRefLike types in generics](#byreflike-generics)
- [Rules for IL rewriters](#rules-for-il-rewriters)
- [Checked user-defined operators](#checked-user-defined-operators)
- [Atomic reads and writes](#atomic-reads-and-writes)
Expand Down Expand Up @@ -1026,6 +1027,31 @@ Changes to signatures:
- Add a bullet point
- Managed pointers which point at null, the address just past the end of an object, or the address where an element just past the end of an array would be stored, are permitted but not dereferenceable.

## <a name="byreflike-generics"></a> ByRefLike types in generics

ByRefLike types, defined in C# with the `ref struct` syntax, represent types that cannot escape to the managed heap and must remain on the stack. It is possible for these types to be used as generic parameters, but in order to improve utility certain affordances are required.

### II.10.1.7
An additional IL keyword, `byreflike`, is introduced to indicate use of ByRefLike types is permitted. This expands the set of permissible types used by this parameters, but limits the potential instructions that can be used on instances of this generic parameter type.

### II.23.1.7
Update the `SpecialConstraintMask` flag value and description, and add a new flag, `AllowByRefLike`.

| Flag | Value | Description |
| --- | ----- | ----------- |
| `SpecialConstraintMask` | `0x3C` | These 4 bits contain one of the following values: |
| ... | ... | ... |
| `AllowByRefLike` | `0x20` | The generic parameter is allowed to be ByRefLike |

### III.2.1
The following case is added as the **third** cases in the "if _thisType_" sequence.

> If _thisType_ is ByRefLike and _thisType_ does not implement _method_ then; a `NotSupportedException` is thrown at the callsite.
The following is added to the paragraph starting with "This last case can only occur when _method_ was defined on `System.Object`, `System.ValueType`, or `System.Enum`".

> The third case can only occur when _method_ was defined on `System.Object` or is a Default Interface Method.
## Rules for IL Rewriters

There are apis such as `System.Runtime.CompilerServices.RuntimeHelpers.CreateSpan<T>(...)` which require that the PE file have a particular structure. In particular, that api requires that the associated RVA of a FieldDef which is used to create a span must be naturally aligned over the data type that `CreateSpan` is instantiated over. There are 2 major concerns.
Expand Down
2 changes: 1 addition & 1 deletion src/coreclr/inc/corinfo.h
Original file line number Diff line number Diff line change
Expand Up @@ -2402,7 +2402,7 @@ class ICorStaticInfo
virtual size_t getClassThreadStaticDynamicInfo (
CORINFO_CLASS_HANDLE cls
) = 0;

virtual bool getStaticBaseAddress(
CORINFO_CLASS_HANDLE cls,
bool isGc,
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -414,10 +414,10 @@ public static unsafe object CheckCastClass(MethodTable* pTargetType, object obj)
[RuntimeExport("RhTypeCast_CheckCastClassSpecial")]
private static unsafe object CheckCastClassSpecial(MethodTable* pTargetType, object obj)
{
Debug.Assert(!pTargetType->IsParameterizedType, "CheckCastClass called with parameterized MethodTable");
Debug.Assert(!pTargetType->IsFunctionPointer, "CheckCastClass called with function pointer MethodTable");
Debug.Assert(!pTargetType->IsInterface, "CheckCastClass called with interface MethodTable");
Debug.Assert(!pTargetType->HasGenericVariance, "CheckCastClass with variant MethodTable");
Debug.Assert(!pTargetType->IsParameterizedType, "CheckCastClassSpecial called with parameterized MethodTable");
Debug.Assert(!pTargetType->IsFunctionPointer, "CheckCastClassSpecial called with function pointer MethodTable");
Debug.Assert(!pTargetType->IsInterface, "CheckCastClassSpecial called with interface MethodTable");
Debug.Assert(!pTargetType->HasGenericVariance, "CheckCastClassSpecial with variant MethodTable");

MethodTable* mt = obj.GetMethodTable();
Debug.Assert(mt != pTargetType, "The check for the trivial cases should be inlined by the JIT");
Expand Down
Loading

0 comments on commit 4ff0041

Please sign in to comment.