|
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
using System.Diagnostics;
using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
namespace System
{
internal static partial class Marvin
{
/// <summary>
/// Compute a Marvin hash and collapse it into a 32-bit hash.
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int ComputeHash32(ReadOnlySpan<byte> data, ulong seed) => ComputeHash32(ref MemoryMarshal.GetReference(data), (uint)data.Length, (uint)seed, (uint)(seed >> 32));
/// <summary>
/// Compute a Marvin hash and collapse it into a 32-bit hash.
/// </summary>
public static int ComputeHash32(ref byte data, uint count, uint p0, uint p1)
{
// Control flow of this method generally flows top-to-bottom, trying to
// minimize the number of branches taken for large (>= 8 bytes, 4 chars) inputs.
// If small inputs (< 8 bytes, 4 chars) are given, this jumps to a "small inputs"
// handler at the end of the method.
if (count < 8)
{
// We can't run the main loop, but we might still have 4 or more bytes available to us.
// If so, jump to the 4 .. 7 bytes logic immediately after the main loop.
if (count >= 4)
{
goto Between4And7BytesRemain;
}
else
{
goto InputTooSmallToEnterMainLoop;
}
}
// Main loop - read 8 bytes at a time.
// The block function is unrolled 2x in this loop.
uint loopCount = count / 8;
Debug.Assert(loopCount > 0, "Shouldn't reach this code path for small inputs.");
do
{
// Most x86 processors have two dispatch ports for reads, so we can read 2x 32-bit
// values in parallel. We opt for this instead of a single 64-bit read since the
// typical use case for Marvin32 is computing String hash codes, and the particular
// layout of String instances means the starting data is never 8-byte aligned when
// running in a 64-bit process.
p0 += Unsafe.ReadUnaligned<uint>(ref data);
uint nextUInt32 = Unsafe.ReadUnaligned<uint>(ref Unsafe.AddByteOffset(ref data, 4));
// One block round for each of the 32-bit integers we just read, 2x rounds total.
Block(ref p0, ref p1);
p0 += nextUInt32;
Block(ref p0, ref p1);
// Bump the data reference pointer and decrement the loop count.
// Decrementing by 1 every time and comparing against zero allows the JIT to produce
// better codegen compared to a standard 'for' loop with an incrementing counter.
// Requires https://github.com/dotnet/runtime/issues/6794 to be addressed first
// before we can realize the full benefits of this.
data = ref Unsafe.AddByteOffset(ref data, 8);
} while (--loopCount > 0);
// n.b. We've not been updating the original 'count' parameter, so its actual value is
// still the original data length. However, we can still rely on its least significant
// 3 bits to tell us how much data remains (0 .. 7 bytes) after the loop above is
// completed.
if ((count & 0b_0100) == 0)
{
goto DoFinalPartialRead;
}
Between4And7BytesRemain:
// If after finishing the main loop we still have 4 or more leftover bytes, or if we had
// 4 .. 7 bytes to begin with and couldn't enter the loop in the first place, we need to
// consume 4 bytes immediately and send them through one round of the block function.
Debug.Assert(count >= 4, "Only should've gotten here if the original count was >= 4.");
p0 += Unsafe.ReadUnaligned<uint>(ref data);
Block(ref p0, ref p1);
DoFinalPartialRead:
// Finally, we have 0 .. 3 bytes leftover. Since we know the original data length was at
// least 4 bytes (smaller lengths are handled at the end of this routine), we can safely
// read the 4 bytes at the end of the buffer without reading past the beginning of the
// original buffer. This necessarily means the data we're about to read will overlap with
// some data we've already processed, but we can handle that below.
Debug.Assert(count >= 4, "Only should've gotten here if the original count was >= 4.");
// Read the last 4 bytes of the buffer.
uint partialResult = Unsafe.ReadUnaligned<uint>(ref Unsafe.Add(ref Unsafe.AddByteOffset(ref data, (nuint)count & 7), -4));
// The 'partialResult' local above contains any data we have yet to read, plus some number
// of bytes which we've already read from the buffer. An example of this is given below
// for little-endian architectures. In this table, AA BB CC are the bytes which we still
// need to consume, and ## are bytes which we want to throw away since we've already
// consumed them as part of a previous read.
//
// (partialResult contains) (we want it to contain)
// count mod 4 = 0 -> [ ## ## ## ## | ] -> 0x####_#### -> 0x0000_0080
// count mod 4 = 1 -> [ ## ## ## ## | AA ] -> 0xAA##_#### -> 0x0000_80AA
// count mod 4 = 2 -> [ ## ## ## ## | AA BB ] -> 0xBBAA_#### -> 0x0080_BBAA
// count mod 4 = 3 -> [ ## ## ## ## | AA BB CC ] -> 0xCCBB_AA## -> 0x80CC_BBAA
count = ~count << 3;
if (BitConverter.IsLittleEndian)
{
partialResult >>= 8; // make some room for the 0x80 byte
partialResult |= 0x8000_0000u; // put the 0x80 byte at the beginning
partialResult >>= (int)count & 0x1F; // shift out all previously consumed bytes
}
else
{
partialResult <<= 8; // make some room for the 0x80 byte
partialResult |= 0x80u; // put the 0x80 byte at the end
partialResult <<= (int)count & 0x1F; // shift out all previously consumed bytes
}
DoFinalRoundsAndReturn:
// Now that we've computed the final partial result, merge it in and run two rounds of
// the block function to finish out the Marvin algorithm.
p0 += partialResult;
Block(ref p0, ref p1);
Block(ref p0, ref p1);
return (int)(p1 ^ p0);
InputTooSmallToEnterMainLoop:
// We had only 0 .. 3 bytes to begin with, so we can't perform any 32-bit reads.
// This means that we're going to be building up the final result right away and
// will only ever run two rounds total of the block function. Let's initialize
// the partial result to "no data".
if (BitConverter.IsLittleEndian)
{
partialResult = 0x80u;
}
else
{
partialResult = 0x80000000u;
}
if ((count & 0b_0001) != 0)
{
// If the buffer is 1 or 3 bytes in length, let's read a single byte now
// and merge it into our partial result. This will result in partialResult
// having one of the two values below, where AA BB CC are the buffer bytes.
//
// (little-endian / big-endian)
// [ AA ] -> 0x0000_80AA / 0xAA80_0000
// [ AA BB CC ] -> 0x0000_80CC / 0xCC80_0000
partialResult = Unsafe.AddByteOffset(ref data, (nuint)count & 2);
if (BitConverter.IsLittleEndian)
{
partialResult |= 0x8000;
}
else
{
partialResult <<= 24;
partialResult |= 0x800000u;
}
}
if ((count & 0b_0010) != 0)
{
// If the buffer is 2 or 3 bytes in length, let's read a single ushort now
// and merge it into the partial result. This will result in partialResult
// having one of the two values below, where AA BB CC are the buffer bytes.
//
// (little-endian / big-endian)
// [ AA BB ] -> 0x0080_BBAA / 0xAABB_8000
// [ AA BB CC ] -> 0x80CC_BBAA / 0xAABB_CC80 (carried over from above)
if (BitConverter.IsLittleEndian)
{
partialResult <<= 16;
partialResult |= (uint)Unsafe.ReadUnaligned<ushort>(ref data);
}
else
{
partialResult |= (uint)Unsafe.ReadUnaligned<ushort>(ref data);
partialResult = BitOperations.RotateLeft(partialResult, 16);
}
}
// Everything is consumed! Go perform the final rounds and return.
goto DoFinalRoundsAndReturn;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void Block(ref uint rp0, ref uint rp1)
{
// Intrinsified in mono interpreter
uint p0 = rp0;
uint p1 = rp1;
p1 ^= p0;
p0 = BitOperations.RotateLeft(p0, 20);
p0 += p1;
p1 = BitOperations.RotateLeft(p1, 9);
p1 ^= p0;
p0 = BitOperations.RotateLeft(p0, 27);
p0 += p1;
p1 = BitOperations.RotateLeft(p1, 19);
rp0 = p0;
rp1 = p1;
}
public static ulong DefaultSeed { get; } = GenerateSeed();
private static unsafe ulong GenerateSeed()
{
ulong seed;
Interop.GetRandomBytes((byte*)&seed, sizeof(ulong));
return seed;
}
}
}
|