|
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
// See the LICENSE file in the project root for more information.
using System;
using System.Diagnostics;
using System.Diagnostics.CodeAnalysis;
using System.Linq;
using System.Threading;
using Microsoft.CodeAnalysis.CSharp.Extensions;
using Microsoft.CodeAnalysis.CSharp.Symbols;
using Microsoft.CodeAnalysis.CSharp.Syntax;
using Microsoft.CodeAnalysis.CSharp.Utilities;
using Microsoft.CodeAnalysis.Operations;
using Microsoft.CodeAnalysis.Shared.Extensions;
using Microsoft.CodeAnalysis.Shared.Utilities;
using Roslyn.Utilities;
namespace Microsoft.CodeAnalysis.CSharp.Simplification.Simplifiers;
/// <summary>
/// By default the cast simplifier operates under several main principles:
/// <list type="number">
/// <item>The final type that a cast-expression was converted to should be the same as the final
/// type that the underlying expression should convert to without the cast. This tells us that
/// the compiler thinks that value should convert to that type implicitly, not just explicitly.</item>
/// <item>Static semantics of the code should remain the same. This means that things like overload
/// resolution of the invocations the casted expression is contained within should not change.</item>
/// <item>Runtime types and values should not observably change. This means that if casting the
/// value would cause a different type to be seen in a <see cref="System.Object.GetType()"/> call,
/// or a different value could be observed at runtime, then it must remain.</item>
/// </list>
///
/// These rules serve as a good foundational intuition about when casts should be kept and when
/// they should be removable. However, they are not entirely complete. There are cases when we
/// can weaken some of the above rules if it would not be observable at runtime. For example,
/// if it can be proven that calling through an interface method would lead to the exact same
/// call at runtime to a specific implementation of that interface method, then it can be legal to
/// remove such a cast as at runtime this would not be observable. This does in effect mean that
/// the emitted IL will be different, but this matches the user expectation that the *end* behavior
/// of their code remains the same.
/// </summary>
internal static class CastSimplifier
{
public static bool IsUnnecessaryCast(ExpressionSyntax cast, SemanticModel semanticModel, CancellationToken cancellationToken)
=> cast switch
{
CastExpressionSyntax castExpression => IsUnnecessaryCast(castExpression, semanticModel, cancellationToken),
BinaryExpressionSyntax binaryExpression => IsUnnecessaryAsCast(binaryExpression, semanticModel, cancellationToken),
_ => false,
};
public static bool IsUnnecessaryAsCast(BinaryExpressionSyntax cast, SemanticModel semanticModel, CancellationToken cancellationToken)
{
return cast.Kind() == SyntaxKind.AsExpression &&
!cast.WalkUpParentheses().ContainsDiagnostics &&
IsCastSafeToRemove(cast, cast.Left, semanticModel, cancellationToken);
}
public static bool IsUnnecessaryCast(CastExpressionSyntax cast, SemanticModel semanticModel, CancellationToken cancellationToken)
{
// Can't remove casts in code that has syntax errors.
if (cast.WalkUpParentheses().ContainsDiagnostics)
return false;
// First handle very special cases where casts are safe to remove, but where we violate
// the general rules of CastSimplifier. Specifically, look for cases where there are multiple
// casts in an expression, which push values out of, but back into the same initial domain,
// and which can be proven to generate the same resultant values with some of the casts
// removed.
//
// This violates the rule that the same set of instructions would be emitted at runtime.
// However, it follows the spirit of the rule that this is not observable, and so removing
// the cast is beneficial to avoid unnecessary work at runtime.
// Special case for: (int)E == 0 case. Enums can always compare against the constant
// 0 without needing a cast.
if (IsEnumCastWithZeroCompare(cast, semanticModel, cancellationToken))
return true;
// Special case for: (E)~(int)e case. Enums don't need to be converted to ints to get bitwise negated.
if (IsRemovableBitwiseEnumNegation(cast, semanticModel, cancellationToken))
return true;
// Special case for converting a method group to object. The compiler issues a warning if the cast is removed:
// warning CS8974: Converting method group 'ToString' to non-delegate type 'object'. Did you intend to invoke the method?
var castExpressionOperation = semanticModel.GetOperation(cast.Expression, cancellationToken);
if (castExpressionOperation is
{
Kind: OperationKind.MethodReference,
Parent.Kind: OperationKind.DelegateCreation,
Parent.Parent: IConversionOperation { Type.SpecialType: SpecialType.System_Object } conversionOperation
})
{
// If we have a double cast, report as unnecessary, e.g:
// (object)(object)MethodGroup
// (Delegate)(object)MethodGroup
// If we have a single object cast, don't report as unnecessary e.g:
// (object)MethodGroup
if (conversionOperation.Parent is IConversionOperation { Type: { } parentConversionType } &&
semanticModel.ClassifyConversion(cast.Expression, parentConversionType).Exists)
{
return true;
}
return false;
}
return IsCastSafeToRemove(cast, cast.Expression, semanticModel, cancellationToken);
}
private static bool IsEnumCastWithZeroCompare(
CastExpressionSyntax castExpression,
SemanticModel semanticModel,
CancellationToken cancellationToken)
{
var leftOrRightChild = castExpression.WalkUpParentheses();
if (leftOrRightChild.Parent is BinaryExpressionSyntax(SyntaxKind.EqualsExpression or SyntaxKind.NotEqualsExpression) binary)
{
var enumType = semanticModel.GetTypeInfo(castExpression.Expression, cancellationToken).Type as INamedTypeSymbol;
var castedType = semanticModel.GetTypeInfo(castExpression.Type, cancellationToken).Type;
if (Equals(enumType?.EnumUnderlyingType, castedType))
{
if (leftOrRightChild == binary.Left && IsConstantZero(binary.Right) ||
leftOrRightChild == binary.Right && IsConstantZero(binary.Left))
{
return true;
}
}
}
return false;
bool IsConstantZero(ExpressionSyntax child)
{
var constantValue = semanticModel.GetConstantValue(child, cancellationToken);
return constantValue.HasValue &&
IntegerUtilities.IsIntegral(constantValue.Value) &&
IntegerUtilities.ToInt64(constantValue.Value) == 0;
}
}
private static bool IsRemovableBitwiseEnumNegation(
CastExpressionSyntax castExpression,
SemanticModel semanticModel,
CancellationToken cancellationToken)
{
// Special case for: (E)~(int)e case or (E?)~(int)e case. Enums don't need to be converted to ints to get
// bitwise negated. The above is equivalent to `~e` as that keeps the same value and the same type.
if (castExpression.WalkUpParentheses().Parent is PrefixUnaryExpressionSyntax(SyntaxKind.BitwiseNotExpression) parent &&
parent.WalkUpParentheses().Parent is CastExpressionSyntax parentCast)
{
var enumType = semanticModel.GetTypeInfo(castExpression.Expression, cancellationToken).Type as INamedTypeSymbol;
var castedType = semanticModel.GetTypeInfo(castExpression.Type, cancellationToken).Type;
if (Equals(enumType?.EnumUnderlyingType, castedType))
{
var parentCastType = semanticModel.GetTypeInfo(parentCast.Type, cancellationToken).Type;
if (Equals(enumType, parentCastType.RemoveNullableIfPresent()))
return true;
}
}
return false;
}
private static bool IsCastSafeToRemove(
ExpressionSyntax castNode, ExpressionSyntax castedExpressionNode,
SemanticModel originalSemanticModel, CancellationToken cancellationToken)
{
#region blocked cases that disqualify this cast from being removed.
// callers should have checked this.
Contract.ThrowIfTrue(castNode.WalkUpParentheses().ContainsDiagnostics);
// Quick syntactic checks we can do before semantic work.
var isDefaultLiteralCast = castedExpressionNode.WalkDownParentheses().IsKind(SyntaxKind.DefaultLiteralExpression);
// Language does not allow `if (x is default)` ever. So if we have `if (x is (Y)default)`
// then we can't remove the cast. This was special cased in the language due to user
// confusion, and so we have to preserve this despite the standard conversion rules
// indicating this should be fine.
if (isDefaultLiteralCast && castNode.WalkUpParentheses().Parent is PatternSyntax or CaseSwitchLabelSyntax)
return false;
#endregion blocked cases
#region allowed cases
// There are cases in the roslyn API where a direct cast does not result in a conversion operation
// (for example, casting a anonymous-method to a delegate type). We have to handle these cases
// specially.
var originalOperation = originalSemanticModel.GetOperation(castNode, cancellationToken);
if (originalOperation is IConversionOperation originalConversionOperation)
{
return IsConversionCastSafeToRemove(
castNode, castedExpressionNode, originalSemanticModel, originalConversionOperation, cancellationToken);
}
if (originalOperation is IDelegateCreationOperation originalDelegateCreationOperation)
{
return IsDelegateCreationCastSafeToRemove(
castNode, castedExpressionNode, originalSemanticModel, originalDelegateCreationOperation, cancellationToken);
}
#endregion allowed cases
return false;
}
private static bool CastRemovalCouldCauseSignExtensionWarning(ExpressionSyntax castSyntax, IConversionOperation conversionOperation)
{
// if we have `... | (T)x` then disallow this cast if we have a widening numeric cast and both T and x are
// signed integers. This can often lead to confusing situations due to sign extension bits getting padded
// to the front of the value. The compiler even warns here in many cases. We don't want to reimplement the
// entire complex compiler algorithm. So just look for the general case and disallow entirely.
//
// Note: it is intentional that this only triggers when both types are integral, and the value that is
// being cast is a signed integer. In other words, the compiler warns both for (long)int, as well as (ulong)int.
if (castSyntax.WalkUpParentheses().GetRequiredParent().Kind() is SyntaxKind.BitwiseOrExpression or SyntaxKind.OrAssignmentExpression)
{
var conversion = conversionOperation.GetConversion();
if (conversion.IsImplicit &&
(conversion.IsNumeric || conversion.IsNullable) &&
conversionOperation.Type.RemoveNullableIfPresent() is var type1 &&
conversionOperation.Operand.Type.RemoveNullableIfPresent() is var type2 &&
type1.IsIntegralType() &&
type2.IsSignedIntegralType())
{
return true;
}
}
return false;
}
private static bool IsDelegateCreationCastSafeToRemove(
ExpressionSyntax castNode, ExpressionSyntax castedExpressionNode,
SemanticModel originalSemanticModel, IDelegateCreationOperation originalDelegateCreationOperation,
CancellationToken cancellationToken)
{
if (originalDelegateCreationOperation.Type?.TypeKind != TypeKind.Delegate)
return false;
// for a cast of an anonymous method to a delegate, we have to make sure that after cast-removal
// that we still have the same.
var (rewrittenSemanticModel, rewrittenExpression) = GetSemanticModelWithCastRemoved(
castNode, castedExpressionNode, originalSemanticModel, cancellationToken);
if (rewrittenSemanticModel is null || rewrittenExpression is null)
return false;
var rewrittenOperation = rewrittenSemanticModel.GetOperation(rewrittenExpression.WalkDownParentheses(), cancellationToken);
if (rewrittenOperation is not (IAnonymousFunctionOperation or IMethodReferenceOperation))
return false;
if (rewrittenOperation.Parent is not IDelegateCreationOperation rewrittenDelegateCreationOperation)
return false;
if (rewrittenDelegateCreationOperation.Type?.TypeKind != TypeKind.Delegate)
return false;
// having to be converting to the same delegate type.
if (!Equals(originalDelegateCreationOperation.Type, rewrittenDelegateCreationOperation.Type))
return false;
// If there are two conversions applied on the same node, ensure both are legal.
if (originalDelegateCreationOperation.Parent is IConversionOperation conversionOperation &&
conversionOperation.Syntax == castNode &&
!IsConversionCastSafeToRemove(castNode, castedExpressionNode, originalSemanticModel, conversionOperation, cancellationToken))
{
return false;
}
return true;
}
private static bool IsConversionCastSafeToRemove(
ExpressionSyntax castNode, ExpressionSyntax castedExpressionNode,
SemanticModel originalSemanticModel, IConversionOperation originalConversionOperation,
CancellationToken cancellationToken)
{
#region blocked cases
// If the conversion doesn't exist then we can't do anything with this as the code isn't
// semantically valid.
var originalConversion = originalConversionOperation.GetConversion();
if (!originalConversion.Exists)
return false;
// A conversion must either not exist, or it must be explicit or implicit. At this point we
// have conversions that will always succeed, but which could have impact on the code by
// changing the types of things (which can affect other things like overload resolution),
// or the runtime values of code. We only want to remove the cast if it will do none of those
// things.
// Explicit conversions are conversions that cannot be proven to always succeed, conversions
// that are known to possibly lose information. As such, we need to preserve this as it
// has necessary runtime behavior that must be kept.
if (IsExplicitCastThatMustBePreserved(originalSemanticModel, castNode, originalConversion, cancellationToken))
return false;
// we are starting with code like `(X)expr` and converting to just `expr`. Post rewrite we need
// to ensure that the final converted-type of `expr` matches the final converted type of `(X)expr`.
var originalConvertedType = originalSemanticModel.GetTypeInfo(castNode.WalkUpParentheses(), cancellationToken).ConvertedType;
if (originalConvertedType is null || originalConvertedType.TypeKind == TypeKind.Error)
return false;
// If removing the cast could cause the compiler to issue a new warning, then we have to preserve it.
if (CastRemovalCouldCauseSignExtensionWarning(castNode, originalConversionOperation))
return false;
// if the expression being casted is the `null` literal, then we can't remove the cast if the final
// converted type isn't known to be a reference type. This can happen with code like:
//
// void Goo<T, S>() where T : class, S
// {
// S y = (T)null;
// }
//
// Effectively, this constrains S to be a reference type (as T could not otherwise derive from it).
// However, such a invariant isn't understood by the compiler. So if the (T) cast is removed it will
// fail as 'null' cannot be converted to an unconstrained generic type.
var isNullLiteralCast = castedExpressionNode.WalkDownParentheses().IsKind(SyntaxKind.NullLiteralExpression);
if (isNullLiteralCast && !originalConvertedType.IsReferenceType && !originalConvertedType.IsNullable())
return false;
// SomeType s = (Action)(() => {}); // Where there's a user defined conversion from Action->SomeType.
//
// This cast is necessary. The language does not allow lambdas to be directly converted to the destination
// type without explicitly stating the intermediary reified delegate type.
var isAnonymousFunctionCast = castedExpressionNode.WalkDownParentheses() is AnonymousFunctionExpressionSyntax;
if (isAnonymousFunctionCast && originalConversion.IsUserDefined)
return false;
// So far, this looks potentially possible to remove. Now, actually do the removal and get the
// semantic model for the rewritten code so we can check it to make sure semantics were preserved.
var (rewrittenSemanticModel, rewrittenExpression) = GetSemanticModelWithCastRemoved(
castNode, castedExpressionNode, originalSemanticModel, cancellationToken);
if (rewrittenSemanticModel is null || rewrittenExpression is null)
return false;
var (rewrittenConvertedType, rewrittenConversion) = GetRewrittenInfo(
castNode, rewrittenExpression,
originalSemanticModel, rewrittenSemanticModel,
originalConversion, originalConvertedType, cancellationToken);
if (rewrittenConvertedType is null || rewrittenConvertedType.TypeKind == TypeKind.Error || !rewrittenConversion.Exists)
return false;
if (CastRemovalWouldCauseUnintendedReferenceComparisonWarning(rewrittenExpression, rewrittenSemanticModel, cancellationToken))
return false;
// The final converted type may be the same even after removing the cast. However, the cast may
// have been necessary to convert the type and/or value in a way that could be observable. For example:
//
// object o1 = (long)expr; // or (long)0
// object o1 = (long?)expr; // or (long?)0
//
// We need to keep the cast so that the stored value stays the right type.
if (originalConversion.IsConstantExpression ||
originalConversion.IsNumeric ||
originalConversion.IsEnumeration ||
originalConversion.IsNullable)
{
if (rewrittenConversion.IsBoxing)
return false;
}
// We have to specially handle formattable string conversions. If we remove them, we may end up with
// a string value instead. For example:
//
// object o2 = (IFormattable)$"";
if (originalConversion.IsInterpolatedString && !rewrittenConversion.IsInterpolatedString)
return false;
// If we have:
//
// public static implicit operator A(string x)
// A x = (string)null;
//
// Then the original code has an implicit user defined conversion in it. We can only remove this
// if the new code would have the same conversion as well.
//
// One special case of this is Span<T> => ReadOnlySpan<T>. This is technically a user-defined-conversion on
// Span<T>, but it can be removed for a collection expression as the compiler knows directly how to make that
// into a ReadOnlySpan
if (originalConversionOperation.Parent is IConversionOperation { Conversion.IsUserDefined: true } originalParentConversionOperation)
{
var originalParentConversion = originalParentConversionOperation.GetConversion();
if (originalParentConversion.IsImplicit)
{
var isAcceptableSpanConversion = originalConversionOperation.Type.IsSpan() && originalParentConversionOperation.Type.IsReadOnlySpan();
if (!isAcceptableSpanConversion)
{
if (!rewrittenConversion.IsUserDefined)
return false;
if (!Equals(originalParentConversion.MethodSymbol, rewrittenConversion.MethodSymbol))
return false;
}
}
}
// Identity fp-casts can actually change the runtime value of the fp number. This can happen because the
// runtime is allowed to perform the operations with wider precision than the actual specified fp-precision.
// i.e. 64-bit doubles can actually be 80 bits at runtime. Even though the language considers this to be an
// identity cast, we don't want to remove these because the user may be depending on that truncation.
if (IsIdentityFloatingPointCastThatMustBePreserved(castNode, castedExpressionNode, originalSemanticModel, cancellationToken))
return false;
// Identity struct casts will make a copy. This copy may need to be kept to preserve semantics that only
// the copy is being manipulated and not the original struct.
if (IsIdentityStructCastThatMustBePreserved(castNode, castedExpressionNode, originalSemanticModel, cancellationToken))
return false;
#endregion blocked cases
#region allowed cases that allow this cast to be removed.
// In code like `((X)y).Z()` the cast to (X) can be removed if the same 'Z' method would be called.
// The rules here can be subtle. For example, if Z is virtual, and (X) is a cast up the inheritance
// hierarchy then this is *normally* ok. HOwever, the language resolve default parameter values
// from the overridden method. So if they differ, we can't actually remove the cast.
//
// Similarly, if (X) is a cast to an interface, and Z is an impl of that interface method, it might
// be possible to remove, but only if y's type is sealed, as otherwise the interface method could be
// reimplemented in a derived type.
//
// Note: this path is fundamentally different from the other forms of cast removal we perform. The
// casts are removed because statically they make no difference to the meaning of the code. Here,
// the code statically changes meaning. However, we can use our knowledge of how the language/runtime
// works to know at *runtime* that the user will get the exact same behavior.
if (castNode.WalkUpParentheses().Parent is MemberAccessExpressionSyntax memberAccessExpression)
{
if (IsComplementaryMemberAccessAfterCastRemoval(
memberAccessExpression, rewrittenExpression, originalSemanticModel, rewrittenSemanticModel, cancellationToken))
{
return true;
}
}
// In code like `((X)y)()` the cast to (X) can be removed if this was an implicit reference conversion
// to a complementary delegate (because of delegate variance) *and* the return type of the delegate
// invoke methods are the same. For example:
//
// Action<object> a = Console.WriteLine;
// ((Action<string>)a)("A");
//
// This is safe as delegate variance ensures that any parameter type has an implicit ref conversion to
// the original delegate type. However, the following would not be safe:
//
// Func<object, string> a = ...;
// var v = ((Func<string, object>)a)("A");
//
// Here the type of 'v' would change to 'object' from 'string'.
//
// Note: this path is fundamentally different from the other forms of cast removal we perform. The
// casts are removed because statically they make no difference to the meaning of the code. Here,
// the code statically changes meaning. However, we can use our knowledge of how the language/runtime
// works to know at *runtime* that the user will get the exact same behavior.
if (castNode.WalkUpParentheses().Parent is InvocationExpressionSyntax invocationExpression)
{
if (IsComplementaryInvocationAfterCastRemoval(
invocationExpression, rewrittenExpression, originalSemanticModel, rewrittenSemanticModel, cancellationToken))
{
return true;
}
}
// If we have an implicit reference conversion in an 'is' expression then we remove the cast. For example
//
// if ((object)someRefType is string)
//
// However if we have:
//
// List<int> list = null;
// if ((object)list is string)
//
// then we don't want to remove the cast as it can cause an error.
if (castNode.WalkUpParentheses().Parent is BinaryExpressionSyntax(SyntaxKind.IsExpression) isExpression &&
originalConversion.IsIdentityOrImplicitReference())
{
var castedExpressionType = originalSemanticModel.GetTypeInfo(castedExpressionNode, cancellationToken).Type;
var isType = originalSemanticModel.GetTypeInfo(isExpression.Right, cancellationToken).Type;
if (castedExpressionType != null && isType != null &&
originalSemanticModel.Compilation.ClassifyConversion(castedExpressionType, isType).Exists)
{
return true;
}
}
// Casts on collection can fundamentally change the runtime representation of the collection. We do not
// want to ever remove them in that case as that's precisely the reason a user may have provided them.
if (originalConversion.IsCollectionExpression || rewrittenConversion.IsCollectionExpression)
{
// Both need to be collection expression conversions, both before and after the cast removal. If not,
// we have no idea what is going on and should not remove.
if (!originalConversion.IsCollectionExpression || !rewrittenConversion.IsCollectionExpression)
return false;
if (IsCollectionExpressionCastThatMustBePreserved(castNode, originalSemanticModel, originalConvertedType, cancellationToken))
return false;
}
// If the types of the expressions are the same, then we can remove safely.
if (originalConvertedType.Equals(rewrittenConvertedType, SymbolEqualityComparer.IncludeNullability))
return true;
// We can safely remove convertion to object in interpolated strings regardless of nullability
if (castNode.IsParentKind(SyntaxKind.Interpolation) && originalConversionOperation.Type?.SpecialType is SpecialType.System_Object)
return true;
// There are cases where the types change but things may still be safe to remove.
// Case1. A value type casted to `object` is safe if it's now getting converted to `dynamic`.
// At runtime `dynamic` is just an `object` as well, and precasting to `object` will end up
// with the same value and type still in the `dynamic` final location.
if (originalConversion.IsBoxing && rewrittenConversion.IsBoxing &&
originalConvertedType.IsReferenceType && rewrittenConvertedType.TypeKind == TypeKind.Dynamic)
{
return true;
}
// There are cases where a cast does have runtime meaning, but can be removed from one location because
// the same effective conversion would happen in the code in a different location. For example:
//
// int? a = b ? (int?)0 : 1
//
// remove this cast will change the meaning of that conditional. It will now produce an int instead of
// an int?. However, we know the same integral value will be produced by the conditional, but will then
// be wrapped with a final conversion back into an int?.
if (IsConditionalCastSafeToRemove(
castNode, originalSemanticModel,
rewrittenExpression, rewrittenSemanticModel, cancellationToken))
{
return true;
}
// Widening a value before bitwise negation produces the same value as bitwise negation
// followed by the same widening. For example:
//
// public static long P(long a, int b)
// => a & ~[|(long)|]b;
if (IsRemovableWideningSignedBitwiseNegation(
castNode, originalConversionOperation,
rewrittenExpression, rewrittenSemanticModel, cancellationToken))
{
return true;
}
// (float?)(int?)2147483647
//
// The inner cast is not necessary here because there is already a lifted nullable conversion
// of the innermost expression to the outer conversion type.
if (IsMultipleImplicitNullableConversion(originalConversionOperation))
return true;
#endregion allowed cases.
return false;
}
private static bool IsCollectionExpressionCastThatMustBePreserved(
ExpressionSyntax castNode,
SemanticModel originalSemanticModel,
ITypeSymbol originalConvertedType,
CancellationToken cancellationToken)
{
// If we can't figure out the type we were casting to, then preserve this cast.
var castedType = originalSemanticModel.GetTypeInfo(castNode, cancellationToken).Type;
if (castedType is null)
return true;
// If the types are exactly the same, like:
//
// int[] x = (int[])[1, 2, 3];
//
// then this is fine to remove.
if (originalConvertedType.Equals(castedType, SymbolEqualityComparer.IncludeNullability))
return false;
// the original and resultant types are *not* the same. This is sometimes ok to remove depending on which
// types it is. Currently, we only support special behavior for named types.
if (castedType is not INamedTypeSymbol namedCastedType ||
originalConvertedType is not INamedTypeSymbol originalNamedConvertedType)
{
return true;
}
if (namedCastedType.TypeArguments.Length != 1 && originalNamedConvertedType.TypeArguments.Length != 1)
return true;
if (!originalNamedConvertedType.TypeArguments[0].Equals(namedCastedType.TypeArguments[0], SymbolEqualityComparer.IncludeNullability))
return true;
// ReadOnlySpan<T> x = (Span<T>)[a, b, c];
//
// unnecessary cast to widened span type. Because it narrows again,
// this is safe to elide as the compiler will go directly to ReadOnlySpan<T> itself.
if (originalConvertedType.IsReadOnlySpan() && castedType.IsSpan())
return false;
// IEnumerable<T> x = (DerivedReadOnlyInterfaceType<T>)[a, b, c]; // also for IReadOnlyCollection and IReadOnlyList
if (originalNamedConvertedType.OriginalDefinition.SpecialType is SpecialType.System_Collections_Generic_IEnumerable_T &&
namedCastedType.OriginalDefinition.SpecialType is SpecialType.System_Collections_Generic_IReadOnlyCollection_T or SpecialType.System_Collections_Generic_IReadOnlyList_T)
{
return false;
}
// ICollection<T> x = (IList<T>)[a, b, c];
if (originalNamedConvertedType.OriginalDefinition.SpecialType is SpecialType.System_Collections_Generic_ICollection_T &&
namedCastedType.OriginalDefinition.SpecialType is SpecialType.System_Collections_Generic_IList_T)
{
return false;
}
// ICollection<T> x = (List<T>)[a, b, c]; // also for IList
if (originalNamedConvertedType.OriginalDefinition.SpecialType is SpecialType.System_Collections_Generic_ICollection_T or SpecialType.System_Collections_Generic_IList_T &&
namedCastedType.OriginalDefinition.Equals(originalSemanticModel.Compilation.ListOfTType()))
{
return false;
}
return true;
}
private static bool IsIdentityStructCastThatMustBePreserved(
ExpressionSyntax castNode, ExpressionSyntax castedExpressionNode, SemanticModel semanticModel, CancellationToken cancellationToken)
{
// Identity struct casts will make a copy. This copy may need to be kept to preserve semantics that only
// the copy is being manipulated and not the original struct.
//
// Note: this is an innacurate heuristic. Generally speaking, practically any member accessed off of a
// struct might mutate it (like accessing .Length on an ImmutableArray). But practically speaking that is
// highly unlikely to actually mutate. To avoid many false negatives from allowing us to simplify pointless
// struct casts, we only look for a very narrow case that just rises up to be potentially problematic.
// Specifically, the invocation of a non-known method on a non-known struct type where neitehr the struct
// nor method are readonly.
var conversion = semanticModel.GetConversion(castedExpressionNode, cancellationToken);
if (!conversion.IsIdentity)
return false;
var castType = semanticModel.GetTypeInfo(castNode, cancellationToken).Type;
if (castType is null)
return false;
// we presume all the built-in types are immutable, so we can skip copying them.
if (castType.IsSpecialType())
return false;
// if it's not a struct, then we're not making a copy and can safely remove the cast.
if (!castType.IsStructType())
return false;
// if the struct is readonly, then we can safely remove the cast as it's not mutable.
if (castType.IsReadOnly)
return false;
// Only have to care if we're actually casting a location (an LVALUE). If we're operating on an rvalue then
// we already have a copy.
var castedSymbol = semanticModel.GetSymbolInfo(castedExpressionNode, cancellationToken).GetAnySymbol();
if (castedSymbol is not IFieldSymbol and not ILocalSymbol and not IParameterSymbol and not IParameterSymbol { RefKind: RefKind.Ref })
return false;
// ok, we have some *potentially* mutable struct. In practice these are rare, but do exist. We'll
// optimistically presume it's ok to remove thist cast *unless* it's the form: `((X)x).SomeMethod()` (where
// SomeMethod is not an override from System.Object and is not readonly itself).
if (castNode.WalkUpParentheses().Parent is not MemberAccessExpressionSyntax { Parent: InvocationExpressionSyntax } memberAccessExpression)
return false;
var memberSymbol = semanticModel.GetSymbolInfo(memberAccessExpression, cancellationToken).GetAnySymbol();
if (memberSymbol is not IMethodSymbol methodSymbol)
return false;
// if it's a readonly method, it's fine to call on the original without copying.
if (methodSymbol.IsReadOnly)
return false;
for (var current = methodSymbol; current != null; current = current.OverriddenMethod)
{
if (current.ContainingType.SpecialType == SpecialType.System_Object)
return false;
}
// Ok, calling some method that could mutate this struct. have to keep this cast.
return true;
}
private static bool IsMultipleImplicitNullableConversion(IConversionOperation originalConversionOperation)
{
// (float?)(int?)2147483647
var innerOriginalConversion = originalConversionOperation.GetConversion();
if (!innerOriginalConversion.IsImplicit || !innerOriginalConversion.IsNullable)
return false;
// if the inner conversion was user defined, we need to keep it as it may have executed user code.
if (innerOriginalConversion.IsUserDefined)
return false;
if (originalConversionOperation.Parent is not IConversionOperation outerOriginalConversionOperation)
return false;
var outerOriginalConversion = outerOriginalConversionOperation.GetConversion();
if (!outerOriginalConversion.IsImplicit || !outerOriginalConversion.IsNullable)
return false;
return true;
}
private static bool IsRemovableWideningSignedBitwiseNegation(
ExpressionSyntax castNode, IConversionOperation originalConversionOperation,
ExpressionSyntax rewrittenExpression, SemanticModel rewrittenSemanticModel,
CancellationToken cancellationToken)
{
// Can potentially remove the cast in:
//
// public static long P(long a, int b)
// => a & ~[|(long)|]b;
//
// We need to have an implicit numeric conversion. Parented by a ~. After removing the cast, we should
// have the same conversion now implicitly on the outside of the `~`.
//
// Similarly, the casted type needs to be the same type we get post rewrite outside the `~`.
//
// Note: this removal only works with signed integers. With unsigned integers the distinction matters.
// Consider ~(ulong)uintVal vs (ulong)~uintVal. the former will extend out the value with 0s, which
// will all be flipped to 1s. The latter will flip any leading 0s to 1s, but will then extend out the
// rest with 1s.
var originalConversion = originalConversionOperation.GetConversion();
if (!originalConversion.IsImplicit || !originalConversion.IsNumeric)
return false;
if (!IsSignedIntegralOrIntPtrType(originalConversionOperation.Type) ||
!IsSignedIntegralOrIntPtrType(originalConversionOperation.Operand.Type))
{
return false;
}
var parent = castNode.WalkUpParentheses().GetRequiredParent();
if (parent is not PrefixUnaryExpressionSyntax(SyntaxKind.BitwiseNotExpression))
return false;
// If we were parented by a bitwise negation before, we must also be afterwards.
var rewrittenBitwiseNotExpression = (PrefixUnaryExpressionSyntax)rewrittenExpression.WalkUpParentheses().GetRequiredParent();
Debug.Assert(rewrittenBitwiseNotExpression.Kind() == SyntaxKind.BitwiseNotExpression);
var rewrittenOperation = rewrittenSemanticModel.GetOperation(rewrittenBitwiseNotExpression, cancellationToken);
if (rewrittenOperation is not IUnaryOperation { OperatorKind: UnaryOperatorKind.BitwiseNegation })
return false;
// Post rewrite we need to have the same conversion outside that `~` that we had inside.
if (rewrittenOperation.Parent is not IConversionOperation rewrittenBitwiseNotConversionOperation)
return false;
var rewrittenBitwiseNotConversion = rewrittenBitwiseNotConversionOperation.GetConversion();
if (originalConversion != rewrittenBitwiseNotConversion)
return false;
// Ensure the types of the cast-inside is the same as the type outside the rewritten `~`.
var originalConvertedType = originalConversionOperation.Type;
var rewrittenBitwiseNotConversionType = rewrittenBitwiseNotConversionOperation.Type;
if (IsNullOrErrorType(originalConvertedType) ||
IsNullOrErrorType(rewrittenBitwiseNotConversionType))
{
return false;
}
if (!originalConvertedType.Equals(rewrittenBitwiseNotConversionType, SymbolEqualityComparer.IncludeNullability))
return false;
return true;
}
private static bool IsSignedIntegralOrIntPtrType(ITypeSymbol? type)
=> type.IsSignedIntegralType() || type?.SpecialType is SpecialType.System_IntPtr;
private static bool IsConditionalCastSafeToRemove(
ExpressionSyntax castNode, SemanticModel originalSemanticModel,
ExpressionSyntax rewrittenExpression, SemanticModel rewrittenSemanticModel, CancellationToken cancellationToken)
{
if (castNode is not CastExpressionSyntax castExpression)
return false;
var parent = castExpression.WalkUpParentheses();
if (parent.Parent is not ConditionalExpressionSyntax originalConditionalExpression)
return false;
// if we were parented by a conditional before, we must be parented by a conditional afterwards.
var rewrittenConditionalExpression = (ConditionalExpressionSyntax)rewrittenExpression.WalkUpParentheses().GetRequiredParent();
if (parent != originalConditionalExpression.WhenFalse && parent != originalConditionalExpression.WhenTrue)
return false;
if (originalSemanticModel.GetOperation(castExpression, cancellationToken) is not IConversionOperation conversionOperation)
return false;
var originalConversion = conversionOperation.GetConversion();
if (!originalConversion.IsNullable && !originalConversion.IsNumeric)
return false;
if (originalConversion.IsNullable)
{
// if we have `a ? (int?)b : default` then we can't remove the nullable cast as it changes the
// meaning of `default`.
if (originalConditionalExpression.WhenTrue.WalkDownParentheses().IsKind(SyntaxKind.DefaultLiteralExpression) ||
originalConditionalExpression.WhenFalse.WalkDownParentheses().IsKind(SyntaxKind.DefaultLiteralExpression))
{
return false;
}
}
var originalCastExpressionTypeInfo = originalSemanticModel.GetTypeInfo(castExpression, cancellationToken);
var originalConditionalTypeInfo = originalSemanticModel.GetTypeInfo(originalConditionalExpression, cancellationToken);
var rewrittenConditionalTypeInfo = rewrittenSemanticModel.GetTypeInfo(rewrittenConditionalExpression, cancellationToken);
if (IsNullOrErrorType(originalCastExpressionTypeInfo) ||
IsNullOrErrorType(originalConditionalTypeInfo) ||
IsNullOrErrorType(rewrittenConditionalTypeInfo))
{
return false;
}
// when we have a ? (T)b : c
//
// then we want the type of the written conditional to be T as well. And we want the final converted
// type of `a ? b : c` to be the same as what `a ? (T)b : c` is converted to.
if (!originalConditionalTypeInfo.ConvertedType!.Equals(rewrittenConditionalTypeInfo.ConvertedType, SymbolEqualityComparer.IncludeNullability))
return false;
var castType = originalSemanticModel.GetTypeInfo(castExpression, cancellationToken).Type;
if (IsNullOrErrorType(castType))
return false;
if (rewrittenSemanticModel.GetOperation(rewrittenConditionalExpression, cancellationToken) is not IConditionalOperation rewrittenConditionalOperation)
return false;
if (castType.Equals(rewrittenConditionalOperation.Type, SymbolEqualityComparer.IncludeNullability))
return true;
if (rewrittenConditionalOperation.Parent is IConversionOperation conditionalParentConversion &&
conditionalParentConversion.GetConversion().IsImplicit &&
castType.Equals(conditionalParentConversion.Type, SymbolEqualityComparer.IncludeNullability))
{
return true;
}
return false;
}
private static bool IsNullOrErrorType(TypeInfo info)
=> IsNullOrErrorType(info.Type) || IsNullOrErrorType(info.ConvertedType);
private static bool IsNullOrErrorType([NotNullWhen(false)] ITypeSymbol? type)
=> type is null || type is IErrorTypeSymbol;
private static bool CastRemovalWouldCauseUnintendedReferenceComparisonWarning(
ExpressionSyntax expression,
SemanticModel semanticModel,
CancellationToken cancellationToken)
{
// Translated from DiagnosticPass.CheckRelationals
var parentBinary = expression.WalkUpParentheses().GetRequiredParent() as BinaryExpressionSyntax;
if (parentBinary != null && parentBinary.Kind() is SyntaxKind.EqualsExpression or SyntaxKind.NotEqualsExpression)
{
var operation = semanticModel.GetOperation(parentBinary, cancellationToken);
if (operation.UnwrapImplicitConversion() is IBinaryOperation binaryOperation)
{
if (binaryOperation.LeftOperand.Type?.SpecialType == SpecialType.System_Object &&
!IsExplicitCast(parentBinary.Left) &&
!IsConstantNull(binaryOperation.LeftOperand) &&
ConvertedHasUserDefinedEquals(binaryOperation.OperatorKind, binaryOperation.RightOperand))
{
return true;
}
else if (binaryOperation.RightOperand.Type?.SpecialType == SpecialType.System_Object &&
!IsExplicitCast(parentBinary.Right) &&
!IsConstantNull(binaryOperation.RightOperand) &&
ConvertedHasUserDefinedEquals(binaryOperation.OperatorKind, binaryOperation.LeftOperand))
{
return true;
}
}
}
return false;
}
private static bool ConvertedHasUserDefinedEquals(BinaryOperatorKind operatorKind, IOperation operation)
{
// translated from DiagnosticPass.ConvertedHasEqual
if (operation is not IConversionOperation conversionOperation)
return false;
if (IsExplicitCast(conversionOperation.Syntax))
return false;
if (conversionOperation.Operand.Type is not INamedTypeSymbol original)
return false;
if (!original.IsReferenceType || original.TypeKind == TypeKind.Interface)
return false;
var opName = operatorKind == BinaryOperatorKind.Equals
? WellKnownMemberNames.EqualityOperatorName
: WellKnownMemberNames.InequalityOperatorName;
for (var type = original; type != null; type = type.BaseType)
{
foreach (var sym in type.GetMembers(opName))
{
if (sym is IMethodSymbol { MethodKind: MethodKind.UserDefinedOperator } op)
{
var parameters = op.GetParameters();
if (parameters.Length == 2 &&
type.Equals(parameters[0].Type) &&
type.Equals(parameters[1].Type))
{
return true;
}
}
}
}
return false;
}
private static bool IsConstantNull(IOperation operation)
=> operation.ConstantValue.HasValue && operation.ConstantValue.Value is null;
private static bool IsExplicitCast(SyntaxNode node)
=> node is ExpressionSyntax expression && expression.WalkDownParentheses().Kind() is SyntaxKind.CastExpression or SyntaxKind.AsExpression;
private static bool IsExplicitCastThatMustBePreserved(
SemanticModel semanticModel,
ExpressionSyntax castOrAsNode,
Conversion conversion,
CancellationToken cancellationToken)
{
if (!conversion.IsExplicit)
return false;
// Some explicit casts are safe to remove as they still will have no runtime impact, (or the compiler would
// insert the implicit cast for it later due to surrounding context).
// Explicit identity casts arise with things like `(string?)""`. In this case, there is no runtime impact,
// just type system impact. This is a candidate for removal, and our later checks will ensure the same
// types remain.
if (conversion.IsIdentity)
return false;
// Explicit nullable casts arise with things like `(int?)0`. These will succeed at runtime, but are potentially
// removable if the language would insert such a cast anyways (for things like `x ? (int?)0 : null`). In C# 9
// and above this will create a legal conditional conversion that implicitly adds that cast.
//
// Note: this does not apply for `as byte?`. This is an explicit as-cast that can produce null values and
// so it should be maintained.
if (conversion.IsNullable && castOrAsNode is CastExpressionSyntax castExpression)
{
var parent = castOrAsNode.WalkUpParentheses();
if (parent.Parent is ConditionalExpressionSyntax conditionalExpression)
{
// If we have `(T?)expr == null` or `null == (T?)expr` then we can potentially remove this cast as
// the lang will implicitly create such a cast with an appropriate type and null.
var (castSide, otherSide) = conditionalExpression.WhenTrue == parent
? (conditionalExpression.WhenTrue, conditionalExpression.WhenFalse)
: (conditionalExpression.WhenFalse, conditionalExpression.WhenTrue);
if (otherSide.WalkDownParentheses().Kind() == SyntaxKind.NullLiteralExpression)
return false;
// if we have `(T?)TExpr == nullableTExpr` then we can also remove this cast as the language will
// insert the same nullable widening cast implicitly.
//
// If we have `(T?)TExpr == TExpr` then we can potentially remove this cast if the caller determines
// that there is an outer contextual cast to `T?` higher up.
var castSideType = semanticModel.GetTypeInfo(castSide, cancellationToken).Type;
var castedExpressionType = semanticModel.GetTypeInfo(castExpression.Expression, cancellationToken).Type;
if (castSideType.IsNullable(out var underlyingType) && Equals(underlyingType, castedExpressionType))
{
var otherSideType = semanticModel.GetTypeInfo(otherSide, cancellationToken).Type;
if (Equals(castSideType, otherSideType) || Equals(underlyingType, otherSideType))
return false;
}
}
}
return true;
}
private static bool IsIdentityFloatingPointCastThatMustBePreserved(
ExpressionSyntax castNode, ExpressionSyntax castedExpressionNode,
SemanticModel semanticModel, CancellationToken cancellationToken)
{
var conversion = semanticModel.GetConversion(castedExpressionNode, cancellationToken);
if (!conversion.IsIdentity)
return false;
var castType = semanticModel.GetTypeInfo(castNode, cancellationToken).Type;
var castedExpressionType = semanticModel.GetTypeInfo(castedExpressionNode, cancellationToken).Type;
// Floating point casts can have subtle runtime behavior, even between the same fp types. For example, a
// cast from float-to-float can still change behavior because it may take a higher precision computation and
// truncate it to 32bits.
//
// Because of this we keep floating point conversions unless we can prove that it's safe. The only safe
// times are when we're loading or storing into a location we know has the same size as the cast size
// (i.e. reading/writing into a field).
if (!IsFloatingPointType(castedExpressionType) ||
!IsFloatingPointType(castType))
{
// wasn't a floating point conversion.
return false;
}
// Identity fp conversion is safe if this is a read from a fp field/array
if (IsFieldOrArrayElement(semanticModel, castedExpressionNode, cancellationToken))
return false;
// Boxing the result will automatically truncate this as well as this must be stored into a real 32bit or
// 64bit location. As such, the explicit cast to truncate to 32/64 isn't necessary. See
// https://github.com/dotnet/roslyn/pull/56932#discussion_r725241921 for more details.
var parentConversion = semanticModel.GetConversion(castNode, cancellationToken);
if (parentConversion.Exists && parentConversion.IsBoxing)
return false;
// It wasn't a read from a fp/field/array. But it might be a write into one.
castNode = castNode.WalkUpParentheses();
if (castNode.Parent is AssignmentExpressionSyntax assignmentExpression &&
assignmentExpression.Right == castNode)
{
// Identity fp conversion is safe if this is a write to a fp field/array
if (IsFieldOrArrayElement(semanticModel, assignmentExpression.Left, cancellationToken))
return false;
}
else if (castNode.Parent is InitializerExpressionSyntax(SyntaxKind.ArrayInitializerExpression) arrayInitializer)
{
// Identity fp conversion is safe if this is in an array initializer.
var typeInfo = semanticModel.GetTypeInfo(arrayInitializer, cancellationToken);
return typeInfo.Type?.Kind == SymbolKind.ArrayType;
}
else if (castNode.Parent is EqualsValueClauseSyntax equalsValue &&
equalsValue.Value == castNode &&
equalsValue.Parent is VariableDeclaratorSyntax variableDeclarator)
{
// Identity fp conversion is safe if this is in a field initializer.
var symbol = semanticModel.GetDeclaredSymbol(variableDeclarator, cancellationToken);
if (symbol?.Kind == SymbolKind.Field)
return false;
}
// We have to preserve this cast.
return true;
}
private static bool IsFloatingPointType(ITypeSymbol? type)
=> type?.SpecialType is SpecialType.System_Double or SpecialType.System_Single;
private static bool IsFieldOrArrayElement(SemanticModel semanticModel, ExpressionSyntax expression, CancellationToken cancellationToken)
{
var operation = semanticModel.GetOperation(expression.WalkDownParentheses(), cancellationToken);
return operation is IFieldReferenceOperation or IArrayElementReferenceOperation;
}
private static bool IntroducedConditionalExpressionConversion(
ExpressionSyntax expression, SemanticModel semanticModel, CancellationToken cancellationToken)
{
for (SyntaxNode? current = expression; current != null; current = current.Parent)
{
var conversion = semanticModel.GetConversion(current, cancellationToken);
if (conversion.IsConditionalExpression)
return true;
}
return false;
}
private static bool IntroducedAmbiguity(
ExpressionSyntax castNode, ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel, SemanticModel rewrittenSemanticModel,
CancellationToken cancellationToken)
{
for (SyntaxNode? currentOld = castNode.WalkUpParentheses().Parent, currentNew = rewrittenExpression.WalkUpParentheses().Parent;
currentOld != null && currentNew != null;
currentOld = currentOld.Parent, currentNew = currentNew.Parent)
{
Debug.Assert(currentOld.Kind() == currentNew.Kind());
var oldSymbolInfo = originalSemanticModel.GetSymbolInfo(currentOld, cancellationToken);
if (oldSymbolInfo.Symbol != null)
{
// if previously we bound to a single symbol, but now we don't, then we introduced an
// error of some sort. Have to bail out immediately and keep the cast.
var newSymbolInfo = rewrittenSemanticModel.GetSymbolInfo(currentNew, cancellationToken);
if (newSymbolInfo.Symbol is null)
return true;
}
if (currentOld is InterpolatedStringExpressionSyntax && currentNew is InterpolatedStringExpressionSyntax)
{
// In the case of interpolations, we need to dive into the operation level to determine if the meaning
// of the the interpolation stayed the same in the case of interpolation handlers.
if (originalSemanticModel.GetOperation(currentOld, cancellationToken) is not IInterpolatedStringOperation oldInterpolationOperation)
return true;
if (rewrittenSemanticModel.GetOperation(currentNew, cancellationToken) is not IInterpolatedStringOperation newInterpolationOperation)
return true;
if (oldInterpolationOperation.Parts.Length != newInterpolationOperation.Parts.Length)
return true;
for (int i = 0, n = oldInterpolationOperation.Parts.Length; i < n; i++)
{
var oldInterpolationPart = oldInterpolationOperation.Parts[i];
var newInterpolationPart = newInterpolationOperation.Parts[i];
if (oldInterpolationPart.Kind != newInterpolationPart.Kind)
return true;
// If we were calling some interpolation AppendFormatted helper, and now we're not, we introduced a problem.
if (oldInterpolationPart is IInterpolatedStringAppendOperation { AppendCall: not IInvalidOperation } &&
newInterpolationPart is IInterpolatedStringAppendOperation { AppendCall: IInvalidOperation })
{
return true;
}
}
}
}
return false;
}
private static bool ChangedOverloadResolution(
ExpressionSyntax castNode, ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel, SemanticModel rewrittenSemanticModel,
CancellationToken cancellationToken)
{
// walk upwards checking overload resolution results. note: we skip until we hit the first argument
// as we don't care about symbol resolution changing when removing a cast in something like `((D)b).X()`
var haveHitArgumentNode = false;
for (SyntaxNode? currentOld = castNode.WalkUpParentheses().Parent, currentNew = rewrittenExpression.WalkUpParentheses().Parent;
currentOld != null && currentNew != null;
currentOld = currentOld.Parent, currentNew = currentNew.Parent)
{
Debug.Assert(currentOld.Kind() == currentNew.Kind());
if (!haveHitArgumentNode && currentOld.Kind() != SyntaxKind.Argument)
continue;
haveHitArgumentNode = true;
var oldSymbolInfo = originalSemanticModel.GetSymbolInfo(currentOld, cancellationToken).Symbol;
var newSymbolInfo = rewrittenSemanticModel.GetSymbolInfo(currentNew, cancellationToken).Symbol;
// ignore local functions. First, we can't test them for equality in speculative situations, but also we
// can't end up with an overload resolution issue for them as they don't have overloads.
if (oldSymbolInfo is IMethodSymbol method &&
method.MethodKind is not (MethodKind.LocalFunction or MethodKind.LambdaMethod) &&
!Equals(oldSymbolInfo, newSymbolInfo))
{
return true;
}
}
return false;
}
private static bool ChangedForEachResolution(
ExpressionSyntax castNode, ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel, SemanticModel rewrittenSemanticModel)
{
for (SyntaxNode? currentOld = castNode.WalkUpParentheses().Parent, currentNew = rewrittenExpression.WalkUpParentheses().Parent;
currentOld != null && currentNew != null;
currentOld = currentOld.Parent, currentNew = currentNew.Parent)
{
Debug.Assert(currentOld.Kind() == currentNew.Kind());
if (currentOld is CommonForEachStatementSyntax oldForEach &&
currentNew is CommonForEachStatementSyntax newForEach)
{
// TODO(cyrusn): Do we need to validate anything else in the foreach infos?
var oldForEachInfo = originalSemanticModel.GetForEachStatementInfo(oldForEach);
var newForEachInfo = rewrittenSemanticModel.GetForEachStatementInfo(newForEach);
var oldConversion = oldForEachInfo.ElementConversion;
var newConversion = newForEachInfo.ElementConversion;
if (oldConversion.IsUserDefined != newConversion.IsUserDefined)
return true;
if (!Equals(oldConversion.MethodSymbol, newConversion.MethodSymbol))
return true;
}
}
return false;
}
private static bool IsComplementaryMemberAccessAfterCastRemoval(
MemberAccessExpressionSyntax memberAccessExpression,
ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel,
SemanticModel rewrittenSemanticModel,
CancellationToken cancellationToken)
{
var originalMemberSymbol = originalSemanticModel.GetSymbolInfo(memberAccessExpression, cancellationToken).Symbol;
if (originalMemberSymbol is null)
return false;
var rewrittenMemberAccessExpression = (MemberAccessExpressionSyntax)rewrittenExpression.WalkUpParentheses().GetRequiredParent();
var rewrittenMemberSymbol = rewrittenSemanticModel.GetSymbolInfo(rewrittenMemberAccessExpression, cancellationToken).Symbol;
if (rewrittenMemberSymbol is null)
return false;
if (originalMemberSymbol.Kind != rewrittenMemberSymbol.Kind)
return false;
// check for: ((X)expr).Invoke(...);
if (IsComplementaryDelegateInvoke(originalMemberSymbol, rewrittenMemberSymbol))
return true;
// Ok, we had two good member symbols before/after the cast removal. In other words we have:
//
// ((X)expr).Y
// (expr).Y
// Next, see if this is a call to an interface method.
if (originalMemberSymbol.ContainingType.TypeKind == TypeKind.Interface)
{
var rewrittenType = rewrittenSemanticModel.GetTypeInfo(rewrittenExpression, cancellationToken).Type;
if (IsNullOrErrorType(rewrittenType))
return false;
// If we don't have a reference type, then it may not be safe to remove the cast. The cast could
// could have been boxing the value and removing that could cause us to operate not on the copy.
//
// Note: intrinsics and enums are also safe as we know they don't have state and thus
// will have the same semantics whether or not they're boxed.
//
// It is also safe if we know the value is already a copy to begin with.
//
// TODO(cyrusn): this may not be true of floating point numbers. Are we sure that it's
// safe to remove an interface cast in that case? Could that cast narrow the precision of
// a wider FP number to a narrower amount (like 80bit FP to 64bit)?
if (!rewrittenType.IsReferenceType &&
!IsIntrinsicOrEnum(rewrittenType) &&
!IsCopy(rewrittenSemanticModel, rewrittenExpression, rewrittenType, cancellationToken))
{
return false;
}
// if we are still calling through to the same interface method, then this is safe to call.
if (Equals(originalMemberSymbol, rewrittenMemberSymbol))
return true;
// Ok, we have a type casted to an interface. It may be safe to remove this interface cast
// if we still call into the implementation of that interface member afterwards. Note: the
// type has to be sealed, otherwise the interface method may have been reimplemented lower
// in the inheritance hierarchy.
//
// However, if this was an object creation expression, then we know the exact type that was
// created, and don't have to worry about subclassing.
var isSealed =
rewrittenType.IsSealed ||
rewrittenType.IsValueType ||
rewrittenType.TypeKind == TypeKind.Array ||
IsIntrinsicOrEnum(rewrittenType) ||
rewrittenExpression.WalkDownParentheses() is ObjectCreationExpressionSyntax;
if (!isSealed)
return false;
// Then look for the current implementation of that interface member.
var rewrittenContainingType = rewrittenMemberSymbol.ContainingType;
var implementationMember = rewrittenContainingType.FindImplementationForInterfaceMember(originalMemberSymbol);
if (implementationMember is null)
return false;
// if that's not the method we're currently calling, then this definitely isn't safe to remove.
return
Equals(implementationMember, rewrittenMemberSymbol) &&
ParameterNamesAndDefaultValuesAndReturnTypesMatch(
memberAccessExpression, originalSemanticModel, originalMemberSymbol, rewrittenMemberSymbol, cancellationToken);
}
// Second, check if this is a virtual call to a different location in the inheritance hierarchy.
// Importantly though, because of covariant return types, we have to make sure the overrides
// agree on the return type, or else this could change the final type of hte expression.
for (var current = rewrittenMemberSymbol; current != null; current = current.GetOverriddenMember())
{
if (Equals(originalMemberSymbol, current))
{
// we're calling into a override of a higher up virtual in the original code.
// This is safe as long as the names of the parameters and all default values
// are the same. This is because the compiler uses the names and default
// values of the overridden member, even though it emits a virtual call to the
// the highest in the inheritance chain.
return ParameterNamesAndDefaultValuesAndReturnTypesMatch(
memberAccessExpression, originalSemanticModel, originalMemberSymbol, rewrittenMemberSymbol, cancellationToken);
}
}
return false;
}
private static bool IsComplementaryInvocationAfterCastRemoval(
InvocationExpressionSyntax memberAccessExpression,
ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel,
SemanticModel rewrittenSemanticModel,
CancellationToken cancellationToken)
{
var originalMemberSymbol = originalSemanticModel.GetSymbolInfo(memberAccessExpression, cancellationToken).Symbol;
if (originalMemberSymbol is null)
return false;
var rewrittenMemberAccessExpression = (InvocationExpressionSyntax)rewrittenExpression.WalkUpParentheses().GetRequiredParent();
var rewrittenMemberSymbol = rewrittenSemanticModel.GetSymbolInfo(rewrittenMemberAccessExpression, cancellationToken).Symbol;
if (rewrittenMemberSymbol is null)
return false;
return IsComplementaryDelegateInvoke(originalMemberSymbol, rewrittenMemberSymbol);
}
private static bool IsComplementaryDelegateInvoke(ISymbol originalMemberSymbol, ISymbol rewrittenMemberSymbol)
{
if (originalMemberSymbol is not IMethodSymbol { MethodKind: MethodKind.DelegateInvoke } originalMethodSymbol ||
rewrittenMemberSymbol is not IMethodSymbol { MethodKind: MethodKind.DelegateInvoke } rewrittenMethodSymbol)
{
return false;
}
// if we're invoking a delegate method, then the removal of the cast is mostly safe (as the
// compiler will only allow implicit reference conversions between variant delegates and
// variant delegates will only allow different implicit reference conversions of their
// parameters and return type.
// However, if the delegate return type differs, then that could change semantics higher
// up, so we must disallow this if they're not the same.
return Equals(originalMethodSymbol.ReturnType, rewrittenMethodSymbol.ReturnType);
}
private static bool IsIntrinsicOrEnum(ITypeSymbol rewrittenType)
=> rewrittenType.IsIntrinsicType() ||
rewrittenType.IsEnumType() ||
rewrittenType.SpecialType == SpecialType.System_Enum;
private static bool IsCopy(
SemanticModel semanticModel,
ExpressionSyntax expression,
ITypeSymbol rewrittenType,
CancellationToken cancellationToken)
{
// Checked by caller first.
Debug.Assert(!rewrittenType.IsReferenceType && !IsIntrinsicOrEnum(rewrittenType));
// Be conservative here. If we can't prove it's not a copy assume it's a copy.
expression = expression.WalkDownParentheses();
var operation = semanticModel.GetOperation(expression, cancellationToken);
if (operation != null)
{
// All operators return a fresh copy. Note: this may need to be revisited if operators
// ever can return byref in the future.
if (operation is IBinaryOperation { OperatorMethod: not null })
return true;
if (operation is IUnaryOperation { OperatorMethod: not null })
return true;
// if we're getting the struct through a non-ref property, then it will make a copy.
if (operation is IPropertyReferenceOperation { Property.RefKind: not RefKind.Ref })
return true;
// if we're getting the struct as the return value of a non-ref method, then it will make a copy.
if (operation is IInvocationOperation { TargetMethod.RefKind: not RefKind.Ref })
return true;
// If we're new'ing up this struct then we have a fresh copy that we can operate on.
if (operation is IObjectCreationOperation)
return true;
}
return false;
}
private static bool ParameterNamesAndDefaultValuesAndReturnTypesMatch(
MemberAccessExpressionSyntax memberAccessExpression, SemanticModel semanticModel,
ISymbol originalMemberSymbol, ISymbol rewrittenMemberSymbol, CancellationToken cancellationToken)
{
var originalMemberType = originalMemberSymbol.GetMemberType();
var rewrittenMemberType = rewrittenMemberSymbol.GetMemberType();
if (!Equals(originalMemberType, rewrittenMemberType))
return false;
// if this member actually invoked, ensure that we end up with the same values for default
// parameters, and that the same names are used. Note: we technically only need to check
// default values for arguments not passed, and we only need to check names for those that
// are passed.
if (memberAccessExpression.GetRequiredParent() is InvocationExpressionSyntax invocationExpression &&
semanticModel.GetOperation(invocationExpression, cancellationToken) is IInvocationOperation invocationOperation)
{
if (originalMemberSymbol is IMethodSymbol originalMethodSymbol &&
rewrittenMemberSymbol is IMethodSymbol rewrittenMethodSymbol)
{
var originalParameters = originalMethodSymbol.Parameters;
var rewrittenParameters = rewrittenMethodSymbol.Parameters;
if (originalParameters.Length != rewrittenParameters.Length)
return false;
for (var i = 0; i < originalParameters.Length; i++)
{
var originalParameter = originalParameters[i];
var rewrittenParameter = rewrittenParameters[i];
var argument = invocationOperation.Arguments.FirstOrDefault(a => Equals(originalParameter, a.Parameter));
var argumentSyntax = argument?.Syntax as ArgumentSyntax;
if (originalParameter.Name != rewrittenParameter.Name &&
argumentSyntax?.NameColon != null)
{
// names are different. this is a problem if the original user code provided a named arg here.
return false;
}
if (originalParameter.HasExplicitDefaultValue &&
rewrittenParameter.HasExplicitDefaultValue &&
!Equals(originalParameter.ExplicitDefaultValue, rewrittenParameter.ExplicitDefaultValue) &&
argumentSyntax == null)
{
// parameter values are different, this is a problem if the original user code did *not* provide
// an argument here.
return false;
}
}
}
}
return true;
}
private static (ITypeSymbol? rewrittenConvertedType, Conversion rewrittenConversion) GetRewrittenInfo(
ExpressionSyntax castNode, ExpressionSyntax rewrittenExpression,
SemanticModel originalSemanticModel, SemanticModel rewrittenSemanticModel,
Conversion originalConversion, ITypeSymbol originalConvertedType,
CancellationToken cancellationToken)
{
if (castNode.WalkUpParentheses().Parent is InterpolationSyntax)
{
// Workaround https://github.com/dotnet/roslyn/issues/56934
// Compiler does not give a conversion inside an interpolation. However, all values in the interpolation
// holes are converted to object.
//
// Note: this may need to be revisited with improved interpolated strings (as they could take
// strongly typed args and could avoid the object boxing).
var convertedType = originalConversion.IsIdentity ? originalConvertedType : originalSemanticModel.Compilation.ObjectType;
return (convertedType, default);
}
var rewrittenConvertedType = rewrittenSemanticModel.GetTypeInfo(rewrittenExpression, cancellationToken).ConvertedType;
var rewrittenConversion = rewrittenSemanticModel.GetConversion(rewrittenExpression, cancellationToken);
return (rewrittenConvertedType, rewrittenConversion);
}
private static (SemanticModel? rewrittenSemanticModel, ExpressionSyntax? rewrittenExpression) GetSemanticModelWithCastRemoved(
ExpressionSyntax castNode,
ExpressionSyntax castedExpressionNode,
SemanticModel originalSemanticModel,
CancellationToken cancellationToken)
{
var analyzer = new SpeculationAnalyzer(castNode, castedExpressionNode, originalSemanticModel, cancellationToken);
var rewrittenExpression = analyzer.ReplacedExpression;
var rewrittenSemanticModel = analyzer.SpeculativeSemanticModel;
// Because of error tolerance in the compiler layer, it's possible for an overload resolution error
// to occur, but all the checks above pass. Specifically, with overload resolution, the binding layer
// will still return results (in lambdas especially) for one of the overloads. For example:
//
// Goo(x => (int)x);
// void Goo(Func<int, object> x)
// Goo(Func<string, object> x)
//
// Here, removing the cast will cause an ambiguity issue. However, the type of 'x' will still appear to
// be an 'int' because of error tolerance. To address this, walk up all containing invocations and
// make sure they're calls to the same methods.
if (IntroducedAmbiguity(castNode, rewrittenExpression, originalSemanticModel, rewrittenSemanticModel, cancellationToken))
return default;
if (ChangedOverloadResolution(castNode, rewrittenExpression, originalSemanticModel, rewrittenSemanticModel, cancellationToken))
return default;
// It's possible that removing a cast in a foreach collection expression will change how the foreach methods
// and conversions resolve. Ensure these stay the same to proceed.
if (ChangedForEachResolution(castNode, rewrittenExpression, originalSemanticModel, rewrittenSemanticModel))
return default;
// Removing a cast may cause a conditional-expression conversion to come into existence. This is
// fine as long as we're in C# 9 or above.
if (originalSemanticModel.Compilation.LanguageVersion() < LanguageVersion.CSharp9 &&
IntroducedConditionalExpressionConversion(rewrittenExpression, rewrittenSemanticModel, cancellationToken))
{
return default;
}
return (rewrittenSemanticModel, rewrittenExpression);
}
}
|