Files
Jan Kotas 64c3405594 Cleanup memory copy helpers (dotnet/coreclr#27307)
- Merge RuntimeImports that has just a few FCalls into Buffer.
- Delete assembly implementation of memcpy for ARM. The implementation was outdated and it was used only on Windows ARM and only for fraction of memcpy callsites.
- Delete unnecessary PInvoke into SysStringLen. We have a managed equivalent on System.Runtime.InteropServices.Marshal.
- Delete large amount of redundant code between SecureString.Windows.cs and SecureString.Unix.cs, and switch SecureString to use Spans
- Fold null check into IsWin32Atom

Signed-off-by: dotnet-bot <[email protected]>
2019-10-22 09:46:29 -04:00

418 lines
17 KiB
C#

// 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.Diagnostics;
using System.Runtime;
using Internal.Runtime.CompilerServices;
#pragma warning disable SA1121 // explicitly using type aliases instead of built-in types
#if BIT64
using nuint = System.UInt64;
#else
using nuint = System.UInt32;
#endif
namespace System
{
internal static partial class SpanHelpers
{
public static unsafe void ClearWithoutReferences(ref byte b, nuint byteLength)
{
if (byteLength == 0)
return;
#if AMD64 || ARM64
// The exact matrix on when ZeroMemory is faster than InitBlockUnaligned is very complex. The factors to consider include
// type of hardware and memory aligment. This threshold was chosen as a good balance accross different configurations.
if (byteLength > 768)
goto PInvoke;
Unsafe.InitBlockUnaligned(ref b, 0, (uint)byteLength);
return;
#else
// TODO: Optimize other platforms to be on par with AMD64 CoreCLR
// Note: It's important that this switch handles lengths at least up to 22.
// See notes below near the main loop for why.
// The switch will be very fast since it can be implemented using a jump
// table in assembly. See http://stackoverflow.com/a/449297/4077294 for more info.
switch (byteLength)
{
case 1:
b = 0;
return;
case 2:
Unsafe.As<byte, short>(ref b) = 0;
return;
case 3:
Unsafe.As<byte, short>(ref b) = 0;
Unsafe.Add<byte>(ref b, 2) = 0;
return;
case 4:
Unsafe.As<byte, int>(ref b) = 0;
return;
case 5:
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.Add<byte>(ref b, 4) = 0;
return;
case 6:
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
return;
case 7:
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.Add<byte>(ref b, 6) = 0;
return;
case 8:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
return;
case 9:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.Add<byte>(ref b, 8) = 0;
return;
case 10:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
return;
case 11:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.Add<byte>(ref b, 10) = 0;
return;
case 12:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
return;
case 13:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.Add<byte>(ref b, 12) = 0;
return;
case 14:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
return;
case 15:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
Unsafe.Add<byte>(ref b, 14) = 0;
return;
case 16:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
return;
case 17:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.Add<byte>(ref b, 16) = 0;
return;
case 18:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 16)) = 0;
return;
case 19:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 16)) = 0;
Unsafe.Add<byte>(ref b, 18) = 0;
return;
case 20:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 16)) = 0;
return;
case 21:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 16)) = 0;
Unsafe.Add<byte>(ref b, 20) = 0;
return;
case 22:
#if BIT64
Unsafe.As<byte, long>(ref b) = 0;
Unsafe.As<byte, long>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
#else
Unsafe.As<byte, int>(ref b) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 12)) = 0;
#endif
Unsafe.As<byte, int>(ref Unsafe.Add<byte>(ref b, 16)) = 0;
Unsafe.As<byte, short>(ref Unsafe.Add<byte>(ref b, 20)) = 0;
return;
}
// P/Invoke into the native version for large lengths
if (byteLength >= 512) goto PInvoke;
nuint i = 0; // byte offset at which we're copying
if (((nuint)Unsafe.AsPointer(ref b) & 3) != 0)
{
if (((nuint)Unsafe.AsPointer(ref b) & 1) != 0)
{
b = 0;
i += 1;
if (((nuint)Unsafe.AsPointer(ref b) & 2) != 0)
goto IntAligned;
}
Unsafe.As<byte, short>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
i += 2;
}
IntAligned:
// On 64-bit IntPtr.Size == 8, so we want to advance to the next 8-aligned address. If
// (int)b % 8 is 0, 5, 6, or 7, we will already have advanced by 0, 3, 2, or 1
// bytes to the next aligned address (respectively), so do nothing. On the other hand,
// if it is 1, 2, 3, or 4 we will want to copy-and-advance another 4 bytes until
// we're aligned.
// The thing 1, 2, 3, and 4 have in common that the others don't is that if you
// subtract one from them, their 3rd lsb will not be set. Hence, the below check.
if ((((nuint)Unsafe.AsPointer(ref b) - 1) & 4) == 0)
{
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
i += 4;
}
nuint end = byteLength - 16;
byteLength -= i; // lower 4 bits of byteLength represent how many bytes are left *after* the unrolled loop
// We know due to the above switch-case that this loop will always run 1 iteration; max
// bytes we clear before checking is 23 (7 to align the pointers, 16 for 1 iteration) so
// the switch handles lengths 0-22.
Debug.Assert(end >= 7 && i <= end);
// This is separated out into a different variable, so the i + 16 addition can be
// performed at the start of the pipeline and the loop condition does not have
// a dependency on the writes.
nuint counter;
do
{
counter = i + 16;
// This loop looks very costly since there appear to be a bunch of temporary values
// being created with the adds, but the jit (for x86 anyways) will convert each of
// these to use memory addressing operands.
// So the only cost is a bit of code size, which is made up for by the fact that
// we save on writes to b.
#if BIT64
Unsafe.As<byte, long>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
Unsafe.As<byte, long>(ref Unsafe.AddByteOffset<byte>(ref b, i + 8)) = 0;
#else
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i + 4)) = 0;
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i + 8)) = 0;
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i + 12)) = 0;
#endif
i = counter;
// See notes above for why this wasn't used instead
// i += 16;
}
while (counter <= end);
if ((byteLength & 8) != 0)
{
#if BIT64
Unsafe.As<byte, long>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
#else
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i + 4)) = 0;
#endif
i += 8;
}
if ((byteLength & 4) != 0)
{
Unsafe.As<byte, int>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
i += 4;
}
if ((byteLength & 2) != 0)
{
Unsafe.As<byte, short>(ref Unsafe.AddByteOffset<byte>(ref b, i)) = 0;
i += 2;
}
if ((byteLength & 1) != 0)
{
Unsafe.AddByteOffset<byte>(ref b, i) = 0;
// We're not using i after this, so not needed
// i += 1;
}
return;
#endif
PInvoke:
Buffer._ZeroMemory(ref b, byteLength);
}
public static unsafe void ClearWithReferences(ref IntPtr ip, nuint pointerSizeLength)
{
Debug.Assert((int)Unsafe.AsPointer(ref ip) % sizeof(IntPtr) == 0, "Should've been aligned on natural word boundary.");
// First write backward 8 natural words at a time.
// Writing backward allows us to get away with only simple modifications to the
// mov instruction's base and index registers between loop iterations.
for (; pointerSizeLength >= 8; pointerSizeLength -= 8)
{
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -1) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -2) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -3) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -4) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -5) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -6) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -7) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -8) = default;
}
Debug.Assert(pointerSizeLength <= 7);
// The logic below works by trying to minimize the number of branches taken for any
// given range of lengths. For example, the lengths [ 4 .. 7 ] are handled by a single
// branch, [ 2 .. 3 ] are handled by a single branch, and [ 1 ] is handled by a single
// branch.
//
// We can write both forward and backward as a perf improvement. For example,
// the lengths [ 4 .. 7 ] can be handled by zeroing out the first four natural
// words and the last 3 natural words. In the best case (length = 7), there are
// no overlapping writes. In the worst case (length = 4), there are three
// overlapping writes near the middle of the buffer. In perf testing, the
// penalty for performing duplicate writes is less expensive than the penalty
// for complex branching.
if (pointerSizeLength >= 4)
{
goto Write4To7;
}
else if (pointerSizeLength >= 2)
{
goto Write2To3;
}
else if (pointerSizeLength > 0)
{
goto Write1;
}
else
{
return; // nothing to write
}
Write4To7:
Debug.Assert(pointerSizeLength >= 4);
// Write first four and last three.
Unsafe.Add(ref ip, 2) = default;
Unsafe.Add(ref ip, 3) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -3) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -2) = default;
Write2To3:
Debug.Assert(pointerSizeLength >= 2);
// Write first two and last one.
Unsafe.Add(ref ip, 1) = default;
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -1) = default;
Write1:
Debug.Assert(pointerSizeLength >= 1);
// Write only element.
ip = default;
}
}
}