/// <summary>
/// This is a second copy of the inner loop of decompressTags used when near the
/// end of the input. The key difference is the reading of the trailer bytes. The
/// fast code does a blind read of the next 4 bytes as an int, and this code
/// assembles the int byte-by-byte to assure that the array is not over run. The
/// reason this code path is separate is the if condition to choose between these
/// two seemingly small differences costs like 10-20% of the throughput. I'm
/// hoping in future' versions of hot-spot this code can be integrated into the
/// main loop but for now it is worth the extra maintenance pain to get the extra
/// 10-20%.
/// </summary>
private static int[] DecompressTagSlow(byte[] input, int ipIndex, byte[] output, int outputLimit, int outputOffset, int opIndex)
{
// read the op code
int opCode = SnappyInternalUtils.LoadByte(input, ipIndex++);
int entry = SnappyInternalUtils.LookupShort(_opLookupTable, opCode);
var trailerBytes = (int)((uint)entry >> 11);
//
// Key difference here
//
int trailer = 0;
switch (trailerBytes)
{
case 4:
trailer = (input[ipIndex + 3] & 0xff) << 24;
goto case 3;
case 3:
trailer |= (input[ipIndex + 2] & 0xff) << 16;
goto case 2;
case 2:
trailer |= (input[ipIndex + 1] & 0xff) << 8;
goto case 1;
case 1:
trailer |= (input[ipIndex] & 0xff);
break;
}
// advance the ipIndex past the op codes
ipIndex += trailerBytes;
int length = entry & 0xff;
if ((opCode & 0x3) == Literal)
{
int literalLength = length + trailer;
CopyLiteral(input, ipIndex, output, opIndex, literalLength);
ipIndex += literalLength;
opIndex += literalLength;
}
else
{
// copyOffset/256 is encoded in bits 8..10. By just fetching
// those bits, we get copyOffset (since the bit-field starts at
// bit 8).
int copyOffset = entry & 0x700;
copyOffset += trailer;
// inline to force hot-spot to keep inline
{
int spaceLeft = outputLimit - opIndex;
int srcIndex = opIndex - copyOffset;
if (srcIndex < outputOffset)
{
throw new CorruptionException("Invalid copy offset for opcode starting at " + (ipIndex - trailerBytes - 1));
}
if (length <= 16 && copyOffset >= 8 && spaceLeft >= 16)
{
// Fast path, used for the majority (70-80%) of dynamic invocations.
SnappyInternalUtils.CopyLong(output, srcIndex, output, opIndex);
SnappyInternalUtils.CopyLong(output, srcIndex + 8, output, opIndex + 8);
}
else if (spaceLeft >= length + MaxIncrementCopyOverflow)
{
IncrementalCopyFastPath(output, srcIndex, opIndex, length);
}
else
{
IncrementalCopy(output, srcIndex, output, opIndex, length);
}
}
opIndex += length;
}
return(new[] { ipIndex, opIndex });
}
private static int DecompressAllTags(
byte[] input,
int inputOffset,
int inputSize,
byte[] output,
int outputOffset)
{
int outputLimit = output.Length;
int ipLimit = inputOffset + inputSize;
int opIndex = outputOffset;
int ipIndex = inputOffset;
while (ipIndex < ipLimit - 5)
{
int opCode = SnappyInternalUtils.LoadByte(input, ipIndex++);
int entry = SnappyInternalUtils.LookupShort(_opLookupTable, opCode);
var trailerBytes = (int)((uint)entry >> 11);
int trailer = ReadTrailer(input, ipIndex, trailerBytes);
// advance the ipIndex past the op codes
ipIndex += (int)((uint)entry >> 11);
int length = entry & 0xff;
if ((opCode & 0x3) == Literal)
{
int literalLength = length + trailer;
CopyLiteral(input, ipIndex, output, opIndex, literalLength);
ipIndex += literalLength;
opIndex += literalLength;
}
else
{
// copyOffset/256 is encoded in bits 8..10. By just fetching
// those bits, we get copyOffset (since the bit-field starts at
// bit 8).
int copyOffset = entry & 0x700;
copyOffset += trailer;
// inline to force hot-spot to keep inline
//
// Equivalent to incrementalCopy (below) except that it can write up to ten extra
// bytes after the end of the copy, and that it is faster.
//
// The main part of this loop is a simple copy of eight bytes at a time until
// we've copied (at least) the requested amount of bytes. However, if op and'
// src are less than eight bytes apart (indicating a repeating pattern of
// length < 8), we first need to expand the pattern in order to get the correct
// results. For instance, if the buffer looks like this, with the eight-byte
// <src> and <op> patterns marked as intervals:
//
// abxxxxxxxxxxxx
// [------] src
// [------] op
//
// a single eight-byte copy from <src> to <op> will repeat the pattern once,
// after which we can move <op> two bytes without moving <src>:
//
// ababxxxxxxxxxx
// [------] src
// [------] op
//
// and repeat the exercise until the two no longer overlap.
//
// This allows us to do very well in the special case of one single byte
// repeated many times, without taking a big hit for more general cases.
//
// The worst case of extra writing past the end of the match occurs when
// op - src == 1 and len == 1; the last copy will read from byte positions
// [0..7] and write to [4..11], whereas it was only supposed to write to
// position 1. Thus, ten excess bytes.
{
int spaceLeft = outputLimit - opIndex;
int srcIndex = opIndex - copyOffset;
if (srcIndex < outputOffset)
{
throw new CorruptionException("Invalid copy offset for opcode starting at " + (ipIndex - trailerBytes - 1));
}
if (length <= 16 && copyOffset >= 8 && spaceLeft >= 16)
{
// Fast path, used for the majority (70-80%) of dynamic invocations.
SnappyInternalUtils.CopyLong(output, srcIndex, output, opIndex);
SnappyInternalUtils.CopyLong(output, srcIndex + 8, output, opIndex + 8);
}
else if (spaceLeft >= length + MaxIncrementCopyOverflow)
{
IncrementalCopyFastPath(output, srcIndex, opIndex, length);
}
else
{
IncrementalCopy(output, srcIndex, output, opIndex, length);
}
}
opIndex += length;
}
}
for (; ipIndex < ipLimit;)
{
int[] result = DecompressTagSlow(input, ipIndex, output, outputLimit, outputOffset, opIndex);
ipIndex = result[0];
opIndex = result[1];
}
return(opIndex - outputOffset);
}
