proposals/csharp-11.0/pattern-match-span-of-char-on-string.md
Span<char> on a constant string[!INCLUDESpecletdisclaimer]
Champion issue: https://github.com/dotnet/csharplang/issues/8640
Permit pattern matching a Span<char> and a ReadOnlySpan<char> on a constant string.
For perfomance, usage of Span<char> and ReadOnlySpan<char> is preferred over string in many scenarios. The framework has added many new APIs to allow you to use ReadOnlySpan<char> in place of a string.
A common operation on strings is to use a switch to test if it is a particular value, and the compiler optimizes such a switch. However there is currently no way to do the same on a ReadOnlySpan<char> efficiently, other than implementing the switch and the optimization manually.
In order to encourage adoption of ReadOnlySpan<char> we allow pattern matching a ReadOnlySpan<char>, on a constant string, thus also allowing it to be used in a switch.
static bool Is123(ReadOnlySpan<char> s)
{
return s is "123";
}
static bool IsABC(Span<char> s)
{
return s switch { "ABC" => true, _ => false };
}
We alter the spec for constant patterns as follows (the proposed addition is shown in bold):
Given a pattern input value
eand a constant patternPwith converted valuev,
- if e has integral type or enum type, or a nullable form of one of those, and v has integral type, the pattern
Pmatches the value e if result of the expressione == vistrue; otherwise- If e is of type
System.Span<char>orSystem.ReadOnlySpan<char>, and c is a constant string, and c does not have a constant value ofnull, then the pattern is considered matching ifSystem.MemoryExtensions.SequenceEqual<char>(e, System.MemoryExtensions.AsSpan(c))returnstrue.- the pattern
Pmatches the value e ifobject.Equals(e, v)returnstrue.
System.Span<T> and System.ReadOnlySpan<T> are matched by name, must be ref structs, and can be defined outside corlib.
System.MemoryExtensions is matched by name and can be defined outside corlib.
The signature of System.MemoryExtensions.SequenceEqual overloads must match:
public static bool SequenceEqual<T>(System.Span<T>, System.ReadOnlySpan<T>)public static bool SequenceEqual<T>(System.ReadOnlySpan<T>, System.ReadOnlySpan<T>)The signature of System.MemoryExtensions.AsSpan must match:
public static System.ReadOnlySpan<char> AsSpan(string)Methods with optional parameters are excluded from consideration.
None
None
Should matching be defined independently from MemoryExtensions.SequenceEqual() etc.?
... the pattern is considered matching if
e.Length == c.Lengthande[i] == c[i]for all characters ine.
Recommendation: Define in terms of MemoryExtensions.SequenceEqual() for performance. If MemoryExtensions is missing, report compile error.
Should matching against (string)null be allowed?
If so, should (string)null subsume "" since MemoryExtensions.AsSpan(null) == MemoryExtensions.AsSpan("")?
static bool IsEmpty(ReadOnlySpan<char> span)
{
return span switch
{
(string)null => true, // ok?
"" => true, // error: unreachable?
_ => false,
};
}
Recommendation: Constant pattern (string)null should be reported as an error.
Should the constant pattern match include a runtime type test of the expression value for Span<char> or ReadOnlySpan<char>?
static bool Is123<T>(Span<T> s)
{
return s is "123"; // test for Span<char>?
}
static bool IsABC<T>(Span<T> s)
{
return s is Span<char> and "ABC"; // ok?
}
static bool IsEmptyString<T>(T t) where T : ref struct
{
return t is ""; // test for ReadOnlySpan<char>, Span<char>, string?
}
Recommendation: No implicit runtime type test for constant pattern. (IsABC<T>() example is allowed because the type test is explicit.)
This recommendation was not implemented. All of the preceding samples produce a compiler error.
Should subsumption consider constant string patterns, list patterns, and Length property pattern?
static int ToNum(ReadOnlySpan<char> s)
{
return s switch
{
{ Length: 0 } => 0,
"" => 1, // error: unreachable?
['A',..] => 2,
"ABC" => 3, // error: unreachable?
_ => 4,
};
}
Recommendation: Same subsumption behavior as used when the expression value is string. (Does that mean no subsumption between constant strings, list patterns, and Length, other than treating [..] as matching any?)