snappy 0.0.15 → 0.0.16

data/{vendor → home/travis/build/miyucy/snappy/vendor}/snappy/snappy-stubs-internal.h

@@ -28,8 +28,8 @@
 //
 // Various stubs for the open-source version of Snappy.
 
-#ifndef UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_
-#define UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_
+#ifndef THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_
+#define THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_
 
 #ifdef HAVE_CONFIG_H
 #include "config.h"
@@ -116,6 +116,15 @@ static const int64 kint64max = static_cast<int64>(0x7FFFFFFFFFFFFFFFLL);
 // sub-architectures.
 //
 // This is a mess, but there's not much we can do about it.
+//
+// To further complicate matters, only LDR instructions (single reads) are
+// allowed to be unaligned, not LDRD (two reads) or LDM (many reads). Unless we
+// explicitly tell the compiler that these accesses can be unaligned, it can and
+// will combine accesses. On armcc, the way to signal this is done by accessing
+// through the type (uint32 __packed *), but GCC has no such attribute
+// (it ignores __attribute__((packed)) on individual variables). However,
+// we can tell it that a _struct_ is unaligned, which has the same effect,
+// so we do that.
 
 #elif defined(__arm__) && \
       !defined(__ARM_ARCH_4__) && \
@@ -131,11 +140,39 @@ static const int64 kint64max = static_cast<int64>(0x7FFFFFFFFFFFFFFFLL);
       !defined(__ARM_ARCH_6ZK__) && \
       !defined(__ARM_ARCH_6T2__)
 
-#define UNALIGNED_LOAD16(_p) (*reinterpret_cast<const uint16 *>(_p))
-#define UNALIGNED_LOAD32(_p) (*reinterpret_cast<const uint32 *>(_p))
+#if __GNUC__
+#define ATTRIBUTE_PACKED __attribute__((__packed__))
+#else
+#define ATTRIBUTE_PACKED
+#endif
 
-#define UNALIGNED_STORE16(_p, _val) (*reinterpret_cast<uint16 *>(_p) = (_val))
-#define UNALIGNED_STORE32(_p, _val) (*reinterpret_cast<uint32 *>(_p) = (_val))
+namespace base {
+namespace internal {
+
+struct Unaligned16Struct {
+  uint16 value;
+  uint8 dummy;  // To make the size non-power-of-two.
+} ATTRIBUTE_PACKED;
+
+struct Unaligned32Struct {
+  uint32 value;
+  uint8 dummy;  // To make the size non-power-of-two.
+} ATTRIBUTE_PACKED;
+
+}  // namespace internal
+}  // namespace base
+
+#define UNALIGNED_LOAD16(_p) \
+    ((reinterpret_cast<const ::snappy::base::internal::Unaligned16Struct *>(_p))->value)
+#define UNALIGNED_LOAD32(_p) \
+    ((reinterpret_cast<const ::snappy::base::internal::Unaligned32Struct *>(_p))->value)
+
+#define UNALIGNED_STORE16(_p, _val) \
+    ((reinterpret_cast< ::snappy::base::internal::Unaligned16Struct *>(_p))->value = \
+         (_val))
+#define UNALIGNED_STORE32(_p, _val) \
+    ((reinterpret_cast< ::snappy::base::internal::Unaligned32Struct *>(_p))->value = \
+         (_val))
 
 // TODO(user): NEON supports unaligned 64-bit loads and stores.
 // See if that would be more efficient on platforms supporting it,
@@ -488,4 +525,4 @@ inline char* string_as_array(string* str) {
 
 }  // namespace snappy
 
-#endif  // UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_
+#endif  // THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_INTERNAL_H_

data/{vendor → home/travis/build/miyucy/snappy/vendor}/snappy/snappy-stubs-public.h.in

@@ -33,8 +33,8 @@
 // which is a public header. Instead, snappy-stubs-public.h is generated by
 // from snappy-stubs-public.h.in at configure time.
 
-#ifndef UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_
-#define UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_
+#ifndef THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_
+#define THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_
 
 #if @ac_cv_have_stdint_h@
 #include <stdint.h>
@@ -80,9 +80,11 @@ typedef unsigned long long uint64;
 
 typedef std::string string;
 
+#ifndef DISALLOW_COPY_AND_ASSIGN
 #define DISALLOW_COPY_AND_ASSIGN(TypeName) \
   TypeName(const TypeName&);               \
   void operator=(const TypeName&)
+#endif
 
 #if !@ac_cv_have_sys_uio_h@
 // Windows does not have an iovec type, yet the concept is universally useful.
@@ -95,4 +97,4 @@ struct iovec {
 
 }  // namespace snappy
 
-#endif  // UTIL_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_
+#endif  // THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_STUBS_PUBLIC_H_

data/{vendor → home/travis/build/miyucy/snappy/vendor}/snappy/snappy-test.cc

@@ -28,13 +28,16 @@
 //
 // Various stubs for the unit tests for the open-source version of Snappy.
 
-#include "snappy-test.h"
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
 
 #ifdef HAVE_WINDOWS_H
-#define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #endif
 
+#include "snappy-test.h"
+
 #include <algorithm>
 
 DEFINE_bool(run_microbenchmarks, true,

data/{vendor → home/travis/build/miyucy/snappy/vendor}/snappy/snappy-test.h

@@ -28,8 +28,8 @@
 //
 // Various stubs for the unit tests for the open-source version of Snappy.
 
-#ifndef UTIL_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_
-#define UTIL_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_
+#ifndef THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_
+#define THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_
 
 #include <iostream>
 #include <string>
@@ -52,7 +52,6 @@
 #endif
 
 #ifdef HAVE_WINDOWS_H
-#define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #endif
 
@@ -132,7 +131,7 @@ namespace File {
 }  // namespace File
 
 namespace file {
-  int Defaults() { }
+  int Defaults() { return 0; }
 
   class DummyStatus {
    public:
@@ -158,6 +157,8 @@ namespace file {
     }
 
     fclose(fp);
+
+    return DummyStatus();
   }
 
   DummyStatus SetContents(const string& filename,
@@ -176,6 +177,8 @@ namespace file {
     }
 
     fclose(fp);
+
+    return DummyStatus();
   }
 }  // namespace file
 
@@ -193,6 +196,7 @@ void Test_Snappy_RandomData();
 void Test_Snappy_FourByteOffset();
 void Test_SnappyCorruption_TruncatedVarint();
 void Test_SnappyCorruption_UnterminatedVarint();
+void Test_SnappyCorruption_OverflowingVarint();
 void Test_Snappy_ReadPastEndOfBuffer();
 void Test_Snappy_FindMatchLength();
 void Test_Snappy_FindMatchLengthRandom();
@@ -497,6 +501,7 @@ static inline int RUN_ALL_TESTS() {
   snappy::Test_Snappy_FourByteOffset();
   snappy::Test_SnappyCorruption_TruncatedVarint();
   snappy::Test_SnappyCorruption_UnterminatedVarint();
+  snappy::Test_SnappyCorruption_OverflowingVarint();
   snappy::Test_Snappy_ReadPastEndOfBuffer();
   snappy::Test_Snappy_FindMatchLength();
   snappy::Test_Snappy_FindMatchLengthRandom();
@@ -544,6 +549,13 @@ class LogMessage {
     PREDICT_TRUE(condition) ? (void)0 : \
     snappy::LogMessageVoidify() & snappy::LogMessageCrash()
 
+#ifdef _MSC_VER
+// ~LogMessageCrash calls abort() and therefore never exits. This is by design
+// so temporarily disable warning C4722.
+#pragma warning(push)
+#pragma warning(disable:4722)
+#endif
+
 class LogMessageCrash : public LogMessage {
  public:
   LogMessageCrash() { }
@@ -553,6 +565,10 @@ class LogMessageCrash : public LogMessage {
   }
 };
 
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif
+
 // This class is used to explicitly ignore values in the conditional
 // logging macros.  This avoids compiler warnings like "value computed
 // is not used" and "statement has no effect".
@@ -572,6 +588,7 @@ class LogMessageVoidify {
 #define CHECK_NE(a, b) CRASH_UNLESS((a) != (b))
 #define CHECK_LT(a, b) CRASH_UNLESS((a) < (b))
 #define CHECK_GT(a, b) CRASH_UNLESS((a) > (b))
+#define CHECK_OK(cond) (cond).CheckSuccess()
 
 }  // namespace
 
@@ -579,4 +596,4 @@ using snappy::CompressFile;
 using snappy::UncompressFile;
 using snappy::MeasureFile;
 
-#endif  // UTIL_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_
+#endif  // THIRD_PARTY_SNAPPY_OPENSOURCE_SNAPPY_TEST_H_

data/{vendor → home/travis/build/miyucy/snappy/vendor}/snappy/snappy.cc

@@ -30,6 +30,9 @@
 #include "snappy-internal.h"
 #include "snappy-sinksource.h"
 
+#if defined(__x86_64__) || defined(_M_X64)
+#include <emmintrin.h>
+#endif
 #include <stdio.h>
 
 #include <algorithm>
@@ -39,6 +42,13 @@
 
 namespace snappy {
 
+using internal::COPY_1_BYTE_OFFSET;
+using internal::COPY_2_BYTE_OFFSET;
+using internal::LITERAL;
+using internal::char_table;
+using internal::kMaximumTagLength;
+using internal::wordmask;
+
 // Any hash function will produce a valid compressed bitstream, but a good
 // hash function reduces the number of collisions and thus yields better
 // compression for compressible input, and more speed for incompressible
@@ -76,79 +86,125 @@ size_t MaxCompressedLength(size_t source_len) {
   return 32 + source_len + source_len/6;
 }
 
-enum {
-  LITERAL = 0,
-  COPY_1_BYTE_OFFSET = 1,  // 3 bit length + 3 bits of offset in opcode
-  COPY_2_BYTE_OFFSET = 2,
-  COPY_4_BYTE_OFFSET = 3
-};
-static const int kMaximumTagLength = 5;  // COPY_4_BYTE_OFFSET plus the actual offset.
-
-// Copy "len" bytes from "src" to "op", one byte at a time.  Used for
-// handling COPY operations where the input and output regions may
-// overlap.  For example, suppose:
-//    src    == "ab"
-//    op     == src + 2
-//    len    == 20
-// After IncrementalCopy(src, op, len), the result will have
-// eleven copies of "ab"
-//    ababababababababababab
-// Note that this does not match the semantics of either memcpy()
-// or memmove().
-static inline void IncrementalCopy(const char* src, char* op, ssize_t len) {
-  assert(len > 0);
-  do {
-    *op++ = *src++;
-  } while (--len > 0);
+namespace {
+
+void UnalignedCopy64(const void* src, void* dst) {
+  memcpy(dst, src, 8);
 }
 
-// Equivalent to IncrementalCopy 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.
+void UnalignedCopy128(const void* src, void* dst) {
+  // TODO(alkis): Remove this when we upgrade to a recent compiler that emits
+  // SSE2 moves for memcpy(dst, src, 16).
+#ifdef __SSE2__
+  __m128i x = _mm_loadu_si128(static_cast<const __m128i*>(src));
+  _mm_storeu_si128(static_cast<__m128i*>(dst), x);
+#else
+  memcpy(dst, src, 16);
+#endif
+}
 
-namespace {
+// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) a byte at a time. Used
+// for handling COPY operations where the input and output regions may overlap.
+// For example, suppose:
+//    src       == "ab"
+//    op        == src + 2
+//    op_limit  == op + 20
+// After IncrementalCopySlow(src, op, op_limit), the result will have eleven
+// copies of "ab"
+//    ababababababababababab
+// Note that this does not match the semantics of either memcpy() or memmove().
+inline char* IncrementalCopySlow(const char* src, char* op,
+                                 char* const op_limit) {
+  while (op < op_limit) {
+    *op++ = *src++;
+  }
+  return op_limit;
+}
 
-const int kMaxIncrementCopyOverflow = 10;
+// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) but faster than
+// IncrementalCopySlow. buf_limit is the address past the end of the writable
+// region of the buffer.
+inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
+                             char* const buf_limit) {
+  // Terminology:
+  //
+  // slop = buf_limit - op
+  // pat  = op - src
+  // len  = limit - op
+  assert(src < op);
+  assert(op_limit <= buf_limit);
+  // NOTE: The compressor always emits 4 <= len <= 64. It is ok to assume that
+  // to optimize this function but we have to also handle these cases in case
+  // the input does not satisfy these conditions.
+
+  size_t pattern_size = op - src;
+  // The cases are split into different branches to allow the branch predictor,
+  // FDO, and static prediction hints to work better. For each input we list the
+  // ratio of invocations that match each condition.
+  //
+  // input        slop < 16   pat < 8  len > 16
+  // ------------------------------------------
+  // html|html4|cp   0%         1.01%    27.73%
+  // urls            0%         0.88%    14.79%
+  // jpg             0%        64.29%     7.14%
+  // pdf             0%         2.56%    58.06%
+  // txt[1-4]        0%         0.23%     0.97%
+  // pb              0%         0.96%    13.88%
+  // bin             0.01%     22.27%    41.17%
+  //
+  // It is very rare that we don't have enough slop for doing block copies. It
+  // is also rare that we need to expand a pattern. Small patterns are common
+  // for incompressible formats and for those we are plenty fast already.
+  // Lengths are normally not greater than 16 but they vary depending on the
+  // input. In general if we always predict len <= 16 it would be an ok
+  // prediction.
+  //
+  // In order to be fast we want a pattern >= 8 bytes and an unrolled loop
+  // copying 2x 8 bytes at a time.
+
+  // Handle the uncommon case where pattern is less than 8 bytes.
+  if (PREDICT_FALSE(pattern_size < 8)) {
+    // Expand pattern to at least 8 bytes. The worse case scenario in terms of
+    // buffer usage is when the pattern is size 3. ^ is the original position
+    // of op. x are irrelevant bytes copied by the last UnalignedCopy64.
+    //
+    // abc
+    // abcabcxxxxx
+    // abcabcabcabcxxxxx
+    //    ^
+    // The last x is 14 bytes after ^.
+    if (PREDICT_TRUE(op <= buf_limit - 14)) {
+      while (pattern_size < 8) {
+        UnalignedCopy64(src, op);
+        op += pattern_size;
+        pattern_size *= 2;
+      }
+      if (PREDICT_TRUE(op >= op_limit)) return op_limit;
+    } else {
+      return IncrementalCopySlow(src, op, op_limit);
+    }
+  }
+  assert(pattern_size >= 8);
 
-inline void IncrementalCopyFastPath(const char* src, char* op, ssize_t len) {
-  while (op - src < 8) {
+  // Copy 2x 8 bytes at a time. Because op - src can be < 16, a single
+  // UnalignedCopy128 might overwrite data in op. UnalignedCopy64 is safe
+  // because expanding the pattern to at least 8 bytes guarantees that
+  // op - src >= 8.
+  while (op <= buf_limit - 16) {
     UnalignedCopy64(src, op);
-    len -= op - src;
-    op += op - src;
+    UnalignedCopy64(src + 8, op + 8);
+    src += 16;
+    op += 16;
+    if (PREDICT_TRUE(op >= op_limit)) return op_limit;
   }
-  while (len > 0) {
+  // We only take this branch if we didn't have enough slop and we can do a
+  // single 8 byte copy.
+  if (PREDICT_FALSE(op <= buf_limit - 8)) {
     UnalignedCopy64(src, op);
     src += 8;
     op += 8;
-    len -= 8;
   }
+  return IncrementalCopySlow(src, op, op_limit);
 }
 
 }  // namespace
@@ -157,26 +213,29 @@ static inline char* EmitLiteral(char* op,
                                 const char* literal,
                                 int len,
                                 bool allow_fast_path) {
-  int n = len - 1;      // Zero-length literals are disallowed
-  if (n < 60) {
+  // The vast majority of copies are below 16 bytes, for which a
+  // call to memcpy is overkill. This fast path can sometimes
+  // copy up to 15 bytes too much, but that is okay in the
+  // main loop, since we have a bit to go on for both sides:
+  //
+  //   - The input will always have kInputMarginBytes = 15 extra
+  //     available bytes, as long as we're in the main loop, and
+  //     if not, allow_fast_path = false.
+  //   - The output will always have 32 spare bytes (see
+  //     MaxCompressedLength).
+  assert(len > 0);      // Zero-length literals are disallowed
+  int n = len - 1;
+  if (allow_fast_path && len <= 16) {
     // Fits in tag byte
     *op++ = LITERAL | (n << 2);
 
-    // The vast majority of copies are below 16 bytes, for which a
-    // call to memcpy is overkill. This fast path can sometimes
-    // copy up to 15 bytes too much, but that is okay in the
-    // main loop, since we have a bit to go on for both sides:
-    //
-    //   - The input will always have kInputMarginBytes = 15 extra
-    //     available bytes, as long as we're in the main loop, and
-    //     if not, allow_fast_path = false.
-    //   - The output will always have 32 spare bytes (see
-    //     MaxCompressedLength).
-    if (allow_fast_path && len <= 16) {
-      UnalignedCopy64(literal, op);
-      UnalignedCopy64(literal + 8, op + 8);
-      return op + len;
-    }
+    UnalignedCopy128(literal, op);
+    return op + len;
+  }
+
+  if (n < 60) {
+    // Fits in tag byte
+    *op++ = LITERAL | (n << 2);
   } else {
     // Encode in upcoming bytes
     char* base = op;
@@ -195,42 +254,54 @@ static inline char* EmitLiteral(char* op,
   return op + len;
 }
 
-static inline char* EmitCopyLessThan64(char* op, size_t offset, int len) {
+static inline char* EmitCopyAtMost64(char* op, size_t offset, size_t len,
+                                     bool len_less_than_12) {
   assert(len <= 64);
   assert(len >= 4);
   assert(offset < 65536);
+  assert(len_less_than_12 == (len < 12));
 
-  if ((len < 12) && (offset < 2048)) {
-    size_t len_minus_4 = len - 4;
-    assert(len_minus_4 < 8);            // Must fit in 3 bits
-    *op++ = COPY_1_BYTE_OFFSET + ((len_minus_4) << 2) + ((offset >> 8) << 5);
+  if (len_less_than_12 && PREDICT_TRUE(offset < 2048)) {
+    // offset fits in 11 bits.  The 3 highest go in the top of the first byte,
+    // and the rest go in the second byte.
+    *op++ = COPY_1_BYTE_OFFSET + ((len - 4) << 2) + ((offset >> 3) & 0xe0);
     *op++ = offset & 0xff;
   } else {
-    *op++ = COPY_2_BYTE_OFFSET + ((len-1) << 2);
-    LittleEndian::Store16(op, offset);
-    op += 2;
+    // Write 4 bytes, though we only care about 3 of them.  The output buffer
+    // is required to have some slack, so the extra byte won't overrun it.
+    uint32 u = COPY_2_BYTE_OFFSET + ((len - 1) << 2) + (offset << 8);
+    LittleEndian::Store32(op, u);
+    op += 3;
   }
   return op;
 }
 
-static inline char* EmitCopy(char* op, size_t offset, int len) {
-  // Emit 64 byte copies but make sure to keep at least four bytes reserved
-  while (len >= 68) {
-    op = EmitCopyLessThan64(op, offset, 64);
-    len -= 64;
-  }
+static inline char* EmitCopy(char* op, size_t offset, size_t len,
+                             bool len_less_than_12) {
+  assert(len_less_than_12 == (len < 12));
+  if (len_less_than_12) {
+    return EmitCopyAtMost64(op, offset, len, true);
+  } else {
+    // A special case for len <= 64 might help, but so far measurements suggest
+    // it's in the noise.
 
-  // Emit an extra 60 byte copy if have too much data to fit in one copy
-  if (len > 64) {
-    op = EmitCopyLessThan64(op, offset, 60);
-    len -= 60;
-  }
+    // Emit 64 byte copies but make sure to keep at least four bytes reserved.
+    while (PREDICT_FALSE(len >= 68)) {
+      op = EmitCopyAtMost64(op, offset, 64, false);
+      len -= 64;
+    }
 
-  // Emit remainder
-  op = EmitCopyLessThan64(op, offset, len);
-  return op;
-}
+    // One or two copies will now finish the job.
+    if (len > 64) {
+      op = EmitCopyAtMost64(op, offset, 60, false);
+      len -= 60;
+    }
 
+    // Emit remainder.
+    op = EmitCopyAtMost64(op, offset, len, len < 12);
+    return op;
+  }
+}
 
 bool GetUncompressedLength(const char* start, size_t n, size_t* result) {
   uint32 v = 0;
@@ -364,9 +435,9 @@ char* CompressFragment(const char* input,
       //
       // Heuristic match skipping: If 32 bytes are scanned with no matches
       // found, start looking only at every other byte. If 32 more bytes are
-      // scanned, look at every third byte, etc.. When a match is found,
-      // immediately go back to looking at every byte. This is a small loss
-      // (~5% performance, ~0.1% density) for compressible data due to more
+      // scanned (or skipped), look at every third byte, etc.. When a match is
+      // found, immediately go back to looking at every byte. This is a small
+      // loss (~5% performance, ~0.1% density) for compressible data due to more
       // bookkeeping, but for non-compressible data (such as JPEG) it's a huge
       // win since the compressor quickly "realizes" the data is incompressible
       // and doesn't bother looking for matches everywhere.
@@ -382,7 +453,8 @@ char* CompressFragment(const char* input,
         ip = next_ip;
         uint32 hash = next_hash;
         assert(hash == Hash(ip, shift));
-        uint32 bytes_between_hash_lookups = skip++ >> 5;
+        uint32 bytes_between_hash_lookups = skip >> 5;
+        skip += bytes_between_hash_lookups;
         next_ip = ip + bytes_between_hash_lookups;
         if (PREDICT_FALSE(next_ip > ip_limit)) {
           goto emit_remainder;
@@ -417,19 +489,21 @@ char* CompressFragment(const char* input,
         // We have a 4-byte match at ip, and no need to emit any
         // "literal bytes" prior to ip.
         const char* base = ip;
-        int matched = 4 + FindMatchLength(candidate + 4, ip + 4, ip_end);
+        std::pair<size_t, bool> p =
+            FindMatchLength(candidate + 4, ip + 4, ip_end);
+        size_t matched = 4 + p.first;
         ip += matched;
         size_t offset = base - candidate;
         assert(0 == memcmp(base, candidate, matched));
-        op = EmitCopy(op, offset, matched);
-        // We could immediately start working at ip now, but to improve
-        // compression we first update table[Hash(ip - 1, ...)].
-        const char* insert_tail = ip - 1;
+        op = EmitCopy(op, offset, matched, p.second);
         next_emit = ip;
         if (PREDICT_FALSE(ip >= ip_limit)) {
           goto emit_remainder;
         }
-        input_bytes = GetEightBytesAt(insert_tail);
+        // We are now looking for a 4-byte match again.  We read
+        // table[Hash(ip, shift)] for that.  To improve compression,
+        // we also update table[Hash(ip - 1, shift)] and table[Hash(ip, shift)].
+        input_bytes = GetEightBytesAt(ip - 1);
         uint32 prev_hash = HashBytes(GetUint32AtOffset(input_bytes, 0), shift);
         table[prev_hash] = ip - base_ip - 1;
         uint32 cur_hash = HashBytes(GetUint32AtOffset(input_bytes, 1), shift);
@@ -493,162 +567,6 @@ char* CompressFragment(const char* input,
 //   bool TryFastAppend(const char* ip, size_t available, size_t length);
 // };
 
-// -----------------------------------------------------------------------
-// Lookup table for decompression code.  Generated by ComputeTable() below.
-// -----------------------------------------------------------------------
-
-// Mapping from i in range [0,4] to a mask to extract the bottom 8*i bits
-static const uint32 wordmask[] = {
-  0u, 0xffu, 0xffffu, 0xffffffu, 0xffffffffu
-};
-
-// Data stored per entry in lookup table:
-//      Range   Bits-used       Description
-//      ------------------------------------
-//      1..64   0..7            Literal/copy length encoded in opcode byte
-//      0..7    8..10           Copy offset encoded in opcode byte / 256
-//      0..4    11..13          Extra bytes after opcode
-//
-// We use eight bits for the length even though 7 would have sufficed
-// because of efficiency reasons:
-//      (1) Extracting a byte is faster than a bit-field
-//      (2) It properly aligns copy offset so we do not need a <<8
-static const uint16 char_table[256] = {
-  0x0001, 0x0804, 0x1001, 0x2001, 0x0002, 0x0805, 0x1002, 0x2002,
-  0x0003, 0x0806, 0x1003, 0x2003, 0x0004, 0x0807, 0x1004, 0x2004,
-  0x0005, 0x0808, 0x1005, 0x2005, 0x0006, 0x0809, 0x1006, 0x2006,
-  0x0007, 0x080a, 0x1007, 0x2007, 0x0008, 0x080b, 0x1008, 0x2008,
-  0x0009, 0x0904, 0x1009, 0x2009, 0x000a, 0x0905, 0x100a, 0x200a,
-  0x000b, 0x0906, 0x100b, 0x200b, 0x000c, 0x0907, 0x100c, 0x200c,
-  0x000d, 0x0908, 0x100d, 0x200d, 0x000e, 0x0909, 0x100e, 0x200e,
-  0x000f, 0x090a, 0x100f, 0x200f, 0x0010, 0x090b, 0x1010, 0x2010,
-  0x0011, 0x0a04, 0x1011, 0x2011, 0x0012, 0x0a05, 0x1012, 0x2012,
-  0x0013, 0x0a06, 0x1013, 0x2013, 0x0014, 0x0a07, 0x1014, 0x2014,
-  0x0015, 0x0a08, 0x1015, 0x2015, 0x0016, 0x0a09, 0x1016, 0x2016,
-  0x0017, 0x0a0a, 0x1017, 0x2017, 0x0018, 0x0a0b, 0x1018, 0x2018,
-  0x0019, 0x0b04, 0x1019, 0x2019, 0x001a, 0x0b05, 0x101a, 0x201a,
-  0x001b, 0x0b06, 0x101b, 0x201b, 0x001c, 0x0b07, 0x101c, 0x201c,
-  0x001d, 0x0b08, 0x101d, 0x201d, 0x001e, 0x0b09, 0x101e, 0x201e,
-  0x001f, 0x0b0a, 0x101f, 0x201f, 0x0020, 0x0b0b, 0x1020, 0x2020,
-  0x0021, 0x0c04, 0x1021, 0x2021, 0x0022, 0x0c05, 0x1022, 0x2022,
-  0x0023, 0x0c06, 0x1023, 0x2023, 0x0024, 0x0c07, 0x1024, 0x2024,
-  0x0025, 0x0c08, 0x1025, 0x2025, 0x0026, 0x0c09, 0x1026, 0x2026,
-  0x0027, 0x0c0a, 0x1027, 0x2027, 0x0028, 0x0c0b, 0x1028, 0x2028,
-  0x0029, 0x0d04, 0x1029, 0x2029, 0x002a, 0x0d05, 0x102a, 0x202a,
-  0x002b, 0x0d06, 0x102b, 0x202b, 0x002c, 0x0d07, 0x102c, 0x202c,
-  0x002d, 0x0d08, 0x102d, 0x202d, 0x002e, 0x0d09, 0x102e, 0x202e,
-  0x002f, 0x0d0a, 0x102f, 0x202f, 0x0030, 0x0d0b, 0x1030, 0x2030,
-  0x0031, 0x0e04, 0x1031, 0x2031, 0x0032, 0x0e05, 0x1032, 0x2032,
-  0x0033, 0x0e06, 0x1033, 0x2033, 0x0034, 0x0e07, 0x1034, 0x2034,
-  0x0035, 0x0e08, 0x1035, 0x2035, 0x0036, 0x0e09, 0x1036, 0x2036,
-  0x0037, 0x0e0a, 0x1037, 0x2037, 0x0038, 0x0e0b, 0x1038, 0x2038,
-  0x0039, 0x0f04, 0x1039, 0x2039, 0x003a, 0x0f05, 0x103a, 0x203a,
-  0x003b, 0x0f06, 0x103b, 0x203b, 0x003c, 0x0f07, 0x103c, 0x203c,
-  0x0801, 0x0f08, 0x103d, 0x203d, 0x1001, 0x0f09, 0x103e, 0x203e,
-  0x1801, 0x0f0a, 0x103f, 0x203f, 0x2001, 0x0f0b, 0x1040, 0x2040
-};
-
-// In debug mode, allow optional computation of the table at startup.
-// Also, check that the decompression table is correct.
-#ifndef NDEBUG
-DEFINE_bool(snappy_dump_decompression_table, false,
-            "If true, we print the decompression table at startup.");
-
-static uint16 MakeEntry(unsigned int extra,
-                        unsigned int len,
-                        unsigned int copy_offset) {
-  // Check that all of the fields fit within the allocated space
-  assert(extra       == (extra & 0x7));          // At most 3 bits
-  assert(copy_offset == (copy_offset & 0x7));    // At most 3 bits
-  assert(len         == (len & 0x7f));           // At most 7 bits
-  return len | (copy_offset << 8) | (extra << 11);
-}
-
-static void ComputeTable() {
-  uint16 dst[256];
-
-  // Place invalid entries in all places to detect missing initialization
-  int assigned = 0;
-  for (int i = 0; i < 256; i++) {
-    dst[i] = 0xffff;
-  }
-
-  // Small LITERAL entries.  We store (len-1) in the top 6 bits.
-  for (unsigned int len = 1; len <= 60; len++) {
-    dst[LITERAL | ((len-1) << 2)] = MakeEntry(0, len, 0);
-    assigned++;
-  }
-
-  // Large LITERAL entries.  We use 60..63 in the high 6 bits to
-  // encode the number of bytes of length info that follow the opcode.
-  for (unsigned int extra_bytes = 1; extra_bytes <= 4; extra_bytes++) {
-    // We set the length field in the lookup table to 1 because extra
-    // bytes encode len-1.
-    dst[LITERAL | ((extra_bytes+59) << 2)] = MakeEntry(extra_bytes, 1, 0);
-    assigned++;
-  }
-
-  // COPY_1_BYTE_OFFSET.
-  //
-  // The tag byte in the compressed data stores len-4 in 3 bits, and
-  // offset/256 in 5 bits.  offset%256 is stored in the next byte.
-  //
-  // This format is used for length in range [4..11] and offset in
-  // range [0..2047]
-  for (unsigned int len = 4; len < 12; len++) {
-    for (unsigned int offset = 0; offset < 2048; offset += 256) {
-      dst[COPY_1_BYTE_OFFSET | ((len-4)<<2) | ((offset>>8)<<5)] =
-        MakeEntry(1, len, offset>>8);
-      assigned++;
-    }
-  }
-
-  // COPY_2_BYTE_OFFSET.
-  // Tag contains len-1 in top 6 bits, and offset in next two bytes.
-  for (unsigned int len = 1; len <= 64; len++) {
-    dst[COPY_2_BYTE_OFFSET | ((len-1)<<2)] = MakeEntry(2, len, 0);
-    assigned++;
-  }
-
-  // COPY_4_BYTE_OFFSET.
-  // Tag contents len-1 in top 6 bits, and offset in next four bytes.
-  for (unsigned int len = 1; len <= 64; len++) {
-    dst[COPY_4_BYTE_OFFSET | ((len-1)<<2)] = MakeEntry(4, len, 0);
-    assigned++;
-  }
-
-  // Check that each entry was initialized exactly once.
-  if (assigned != 256) {
-    fprintf(stderr, "ComputeTable: assigned only %d of 256\n", assigned);
-    abort();
-  }
-  for (int i = 0; i < 256; i++) {
-    if (dst[i] == 0xffff) {
-      fprintf(stderr, "ComputeTable: did not assign byte %d\n", i);
-      abort();
-    }
-  }
-
-  if (FLAGS_snappy_dump_decompression_table) {
-    printf("static const uint16 char_table[256] = {\n  ");
-    for (int i = 0; i < 256; i++) {
-      printf("0x%04x%s",
-             dst[i],
-             ((i == 255) ? "\n" : (((i%8) == 7) ? ",\n  " : ", ")));
-    }
-    printf("};\n");
-  }
-
-  // Check that computed table matched recorded table
-  for (int i = 0; i < 256; i++) {
-    if (dst[i] != char_table[i]) {
-      fprintf(stderr, "ComputeTable: byte %d: computed (%x), expect (%x)\n",
-              i, static_cast<int>(dst[i]), static_cast<int>(char_table[i]));
-      abort();
-    }
-  }
-}
-#endif /* !NDEBUG */
 
 // Helper class for decompression
 class SnappyDecompressor {
@@ -701,7 +619,9 @@ class SnappyDecompressor {
       if (n == 0) return false;
       const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip));
       reader_->Skip(1);
-      *result |= static_cast<uint32>(c & 0x7f) << shift;
+      uint32 val = c & 0x7f;
+      if (((val << shift) >> shift) != val) return false;
+      *result |= val << shift;
       if (c < 128) {
         break;
       }
@@ -715,6 +635,10 @@ class SnappyDecompressor {
   template <class Writer>
   void DecompressAllTags(Writer* writer) {
     const char* ip = ip_;
+    // For position-independent executables, accessing global arrays can be
+    // slow.  Move wordmask array onto the stack to mitigate this.
+    uint32 wordmask[sizeof(internal::wordmask)/sizeof(uint32)];
+    memcpy(wordmask, internal::wordmask, sizeof(wordmask));
 
     // We could have put this refill fragment only at the beginning of the loop.
     // However, duplicating it at the end of each branch gives the compiler more
@@ -731,7 +655,19 @@ class SnappyDecompressor {
     for ( ;; ) {
       const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip++));
 
-      if ((c & 0x3) == LITERAL) {
+      // Ratio of iterations that have LITERAL vs non-LITERAL for different
+      // inputs.
+      //
+      // input          LITERAL  NON_LITERAL
+      // -----------------------------------
+      // html|html4|cp   23%        77%
+      // urls            36%        64%
+      // jpg             47%        53%
+      // pdf             19%        81%
+      // txt[1-4]        25%        75%
+      // pb              24%        76%
+      // bin             24%        76%
+      if (PREDICT_FALSE((c & 0x3) == LITERAL)) {
         size_t literal_length = (c >> 2) + 1u;
         if (writer->TryFastAppend(ip, ip_limit_ - ip, literal_length)) {
           assert(literal_length < 61);
@@ -767,15 +703,15 @@ class SnappyDecompressor {
         ip += literal_length;
         MAYBE_REFILL();
       } else {
-        const uint32 entry = char_table[c];
-        const uint32 trailer = LittleEndian::Load32(ip) & wordmask[entry >> 11];
-        const uint32 length = entry & 0xff;
+        const size_t entry = char_table[c];
+        const size_t trailer = LittleEndian::Load32(ip) & wordmask[entry >> 11];
+        const size_t length = entry & 0xff;
         ip += entry >> 11;
 
         // copy_offset/256 is encoded in bits 8..10.  By just fetching
         // those bits, we get copy_offset (since the bit-field starts at
         // bit 8).
-        const uint32 copy_offset = entry & 0x700;
+        const size_t copy_offset = entry & 0x700;
         if (!writer->AppendFromSelf(copy_offset + trailer, length)) {
           return;
         }
@@ -795,10 +731,8 @@ bool SnappyDecompressor::RefillTag() {
     size_t n;
     ip = reader_->Peek(&n);
     peeked_ = n;
-    if (n == 0) {
-      eof_ = true;
-      return false;
-    }
+    eof_ = (n == 0);
+    if (eof_) return false;
     ip_limit_ = ip + n;
   }
 
@@ -863,6 +797,7 @@ static bool InternalUncompressAllTags(SnappyDecompressor* decompressor,
 
   // Process the entire input
   decompressor->DecompressAllTags(writer);
+  writer->Flush();
   return (decompressor->eof() && writer->CheckLength());
 }
 
@@ -965,7 +900,7 @@ class SnappyIOVecWriter {
   const size_t output_iov_count_;
 
   // We are currently writing into output_iov_[curr_iov_index_].
-  int curr_iov_index_;
+  size_t curr_iov_index_;
 
   // Bytes written to output_iov_[curr_iov_index_] so far.
   size_t curr_iov_written_;
@@ -976,7 +911,7 @@ class SnappyIOVecWriter {
   // Maximum number of bytes that will be decompressed into output_iov_.
   size_t output_limit_;
 
-  inline char* GetIOVecPointer(int index, size_t offset) {
+  inline char* GetIOVecPointer(size_t index, size_t offset) {
     return reinterpret_cast<char*>(output_iov_[index].iov_base) +
         offset;
   }
@@ -1037,8 +972,7 @@ class SnappyIOVecWriter {
         output_iov_[curr_iov_index_].iov_len - curr_iov_written_ >= 16) {
       // Fast path, used for the majority (about 95%) of invocations.
       char* ptr = GetIOVecPointer(curr_iov_index_, curr_iov_written_);
-      UnalignedCopy64(ip, ptr);
-      UnalignedCopy64(ip + 8, ptr + 8);
+      UnalignedCopy128(ip, ptr);
       curr_iov_written_ += len;
       total_written_ += len;
       return true;
@@ -1057,7 +991,7 @@ class SnappyIOVecWriter {
     }
 
     // Locate the iovec from which we need to start the copy.
-    int from_iov_index = curr_iov_index_;
+    size_t from_iov_index = curr_iov_index_;
     size_t from_iov_offset = curr_iov_written_;
     while (offset > 0) {
       if (from_iov_offset >= offset) {
@@ -1066,8 +1000,8 @@ class SnappyIOVecWriter {
       }
 
       offset -= from_iov_offset;
+      assert(from_iov_index > 0);
       --from_iov_index;
-      assert(from_iov_index >= 0);
       from_iov_offset = output_iov_[from_iov_index].iov_len;
     }
 
@@ -1102,9 +1036,10 @@ class SnappyIOVecWriter {
         if (to_copy > len) {
           to_copy = len;
         }
-        IncrementalCopy(GetIOVecPointer(from_iov_index, from_iov_offset),
-                        GetIOVecPointer(curr_iov_index_, curr_iov_written_),
-                        to_copy);
+        IncrementalCopySlow(
+            GetIOVecPointer(from_iov_index, from_iov_offset),
+            GetIOVecPointer(curr_iov_index_, curr_iov_written_),
+            GetIOVecPointer(curr_iov_index_, curr_iov_written_) + to_copy);
         curr_iov_written_ += to_copy;
         from_iov_offset += to_copy;
         total_written_ += to_copy;
@@ -1115,6 +1050,7 @@ class SnappyIOVecWriter {
     return true;
   }
 
+  inline void Flush() {}
 };
 
 bool RawUncompressToIOVec(const char* compressed, size_t compressed_length,
@@ -1145,7 +1081,8 @@ class SnappyArrayWriter {
  public:
   inline explicit SnappyArrayWriter(char* dst)
       : base_(dst),
-        op_(dst) {
+        op_(dst),
+        op_limit_(dst) {
   }
 
   inline void SetExpectedLength(size_t len) {
@@ -1172,8 +1109,7 @@ class SnappyArrayWriter {
     const size_t space_left = op_limit_ - op;
     if (len <= 16 && available >= 16 + kMaximumTagLength && space_left >= 16) {
       // Fast path, used for the majority (about 95%) of invocations.
-      UnalignedCopy64(ip, op);
-      UnalignedCopy64(ip + 8, op + 8);
+      UnalignedCopy128(ip, op);
       op_ = op + len;
       return true;
     } else {
@@ -1182,8 +1118,7 @@ class SnappyArrayWriter {
   }
 
   inline bool AppendFromSelf(size_t offset, size_t len) {
-    char* op = op_;
-    const size_t space_left = op_limit_ - op;
+    char* const op_end = op_ + len;
 
     // Check if we try to append from before the start of the buffer.
     // Normally this would just be a check for "produced < offset",
@@ -1192,29 +1127,16 @@ class SnappyArrayWriter {
     // to a very big number. This is convenient, as offset==0 is another
     // invalid case that we also want to catch, so that we do not go
     // into an infinite loop.
-    assert(op >= base_);
-    size_t produced = op - base_;
-    if (produced <= offset - 1u) {
-      return false;
-    }
-    if (len <= 16 && offset >= 8 && space_left >= 16) {
-      // Fast path, used for the majority (70-80%) of dynamic invocations.
-      UnalignedCopy64(op - offset, op);
-      UnalignedCopy64(op - offset + 8, op + 8);
-    } else {
-      if (space_left >= len + kMaxIncrementCopyOverflow) {
-        IncrementalCopyFastPath(op - offset, op, len);
-      } else {
-        if (space_left < len) {
-          return false;
-        }
-        IncrementalCopy(op - offset, op, len);
-      }
-    }
+    if (Produced() <= offset - 1u || op_end > op_limit_) return false;
+    op_ = IncrementalCopy(op_ - offset, op_, op_end, op_limit_);
 
-    op_ = op + len;
     return true;
   }
+  inline size_t Produced() const {
+    assert(op_ >= base_);
+    return op_ - base_;
+  }
+  inline void Flush() {}
 };
 
 bool RawUncompress(const char* compressed, size_t n, char* uncompressed) {
@@ -1241,7 +1163,6 @@ bool Uncompress(const char* compressed, size_t n, string* uncompressed) {
   return RawUncompress(compressed, n, string_as_array(uncompressed));
 }
 
-
 // A Writer that drops everything on the floor and just does validation
 class SnappyDecompressionValidator {
  private:
@@ -1249,7 +1170,7 @@ class SnappyDecompressionValidator {
   size_t produced_;
 
  public:
-  inline SnappyDecompressionValidator() : produced_(0) { }
+  inline SnappyDecompressionValidator() : expected_(0), produced_(0) { }
   inline void SetExpectedLength(size_t len) {
     expected_ = len;
   }
@@ -1270,6 +1191,7 @@ class SnappyDecompressionValidator {
     produced_ += len;
     return produced_ <= expected_;
   }
+  inline void Flush() {}
 };
 
 bool IsValidCompressedBuffer(const char* compressed, size_t n) {
@@ -1278,6 +1200,11 @@ bool IsValidCompressedBuffer(const char* compressed, size_t n) {
   return InternalUncompress(&reader, &writer);
 }
 
+bool IsValidCompressed(Source* compressed) {
+  SnappyDecompressionValidator writer;
+  return InternalUncompress(compressed, &writer);
+}
+
 void RawCompress(const char* input,
                  size_t input_length,
                  char* compressed,
@@ -1292,7 +1219,7 @@ void RawCompress(const char* input,
 
 size_t Compress(const char* input, size_t input_length, string* compressed) {
   // Pre-grow the buffer to the max length of the compressed output
-  compressed->resize(MaxCompressedLength(input_length));
+  STLStringResizeUninitialized(compressed, MaxCompressedLength(input_length));
 
   size_t compressed_length;
   RawCompress(input, input_length, string_as_array(compressed),
@@ -1301,6 +1228,237 @@ size_t Compress(const char* input, size_t input_length, string* compressed) {
   return compressed_length;
 }
 
+// -----------------------------------------------------------------------
+// Sink interface
+// -----------------------------------------------------------------------
 
-} // end namespace snappy
+// A type that decompresses into a Sink. The template parameter
+// Allocator must export one method "char* Allocate(int size);", which
+// allocates a buffer of "size" and appends that to the destination.
+template <typename Allocator>
+class SnappyScatteredWriter {
+  Allocator allocator_;
+
+  // We need random access into the data generated so far.  Therefore
+  // we keep track of all of the generated data as an array of blocks.
+  // All of the blocks except the last have length kBlockSize.
+  std::vector<char*> blocks_;
+  size_t expected_;
+
+  // Total size of all fully generated blocks so far
+  size_t full_size_;
+
+  // Pointer into current output block
+  char* op_base_;       // Base of output block
+  char* op_ptr_;        // Pointer to next unfilled byte in block
+  char* op_limit_;      // Pointer just past block
+
+  inline size_t Size() const {
+    return full_size_ + (op_ptr_ - op_base_);
+  }
+
+  bool SlowAppend(const char* ip, size_t len);
+  bool SlowAppendFromSelf(size_t offset, size_t len);
+
+ public:
+  inline explicit SnappyScatteredWriter(const Allocator& allocator)
+      : allocator_(allocator),
+        full_size_(0),
+        op_base_(NULL),
+        op_ptr_(NULL),
+        op_limit_(NULL) {
+  }
+
+  inline void SetExpectedLength(size_t len) {
+    assert(blocks_.empty());
+    expected_ = len;
+  }
+
+  inline bool CheckLength() const {
+    return Size() == expected_;
+  }
+
+  // Return the number of bytes actually uncompressed so far
+  inline size_t Produced() const {
+    return Size();
+  }
+
+  inline bool Append(const char* ip, size_t len) {
+    size_t avail = op_limit_ - op_ptr_;
+    if (len <= avail) {
+      // Fast path
+      memcpy(op_ptr_, ip, len);
+      op_ptr_ += len;
+      return true;
+    } else {
+      return SlowAppend(ip, len);
+    }
+  }
+
+  inline bool TryFastAppend(const char* ip, size_t available, size_t length) {
+    char* op = op_ptr_;
+    const int space_left = op_limit_ - op;
+    if (length <= 16 && available >= 16 + kMaximumTagLength &&
+        space_left >= 16) {
+      // Fast path, used for the majority (about 95%) of invocations.
+      UnalignedCopy128(ip, op);
+      op_ptr_ = op + length;
+      return true;
+    } else {
+      return false;
+    }
+  }
 
+  inline bool AppendFromSelf(size_t offset, size_t len) {
+    char* const op_end = op_ptr_ + len;
+    // See SnappyArrayWriter::AppendFromSelf for an explanation of
+    // the "offset - 1u" trick.
+    if (PREDICT_TRUE(offset - 1u < op_ptr_ - op_base_ && op_end <= op_limit_)) {
+      // Fast path: src and dst in current block.
+      op_ptr_ = IncrementalCopy(op_ptr_ - offset, op_ptr_, op_end, op_limit_);
+      return true;
+    }
+    return SlowAppendFromSelf(offset, len);
+  }
+
+  // Called at the end of the decompress. We ask the allocator
+  // write all blocks to the sink.
+  inline void Flush() { allocator_.Flush(Produced()); }
+};
+
+template<typename Allocator>
+bool SnappyScatteredWriter<Allocator>::SlowAppend(const char* ip, size_t len) {
+  size_t avail = op_limit_ - op_ptr_;
+  while (len > avail) {
+    // Completely fill this block
+    memcpy(op_ptr_, ip, avail);
+    op_ptr_ += avail;
+    assert(op_limit_ - op_ptr_ == 0);
+    full_size_ += (op_ptr_ - op_base_);
+    len -= avail;
+    ip += avail;
+
+    // Bounds check
+    if (full_size_ + len > expected_) {
+      return false;
+    }
+
+    // Make new block
+    size_t bsize = min<size_t>(kBlockSize, expected_ - full_size_);
+    op_base_ = allocator_.Allocate(bsize);
+    op_ptr_ = op_base_;
+    op_limit_ = op_base_ + bsize;
+    blocks_.push_back(op_base_);
+    avail = bsize;
+  }
+
+  memcpy(op_ptr_, ip, len);
+  op_ptr_ += len;
+  return true;
+}
+
+template<typename Allocator>
+bool SnappyScatteredWriter<Allocator>::SlowAppendFromSelf(size_t offset,
+                                                         size_t len) {
+  // Overflow check
+  // See SnappyArrayWriter::AppendFromSelf for an explanation of
+  // the "offset - 1u" trick.
+  const size_t cur = Size();
+  if (offset - 1u >= cur) return false;
+  if (expected_ - cur < len) return false;
+
+  // Currently we shouldn't ever hit this path because Compress() chops the
+  // input into blocks and does not create cross-block copies. However, it is
+  // nice if we do not rely on that, since we can get better compression if we
+  // allow cross-block copies and thus might want to change the compressor in
+  // the future.
+  size_t src = cur - offset;
+  while (len-- > 0) {
+    char c = blocks_[src >> kBlockLog][src & (kBlockSize-1)];
+    Append(&c, 1);
+    src++;
+  }
+  return true;
+}
+
+class SnappySinkAllocator {
+ public:
+  explicit SnappySinkAllocator(Sink* dest): dest_(dest) {}
+  ~SnappySinkAllocator() {}
+
+  char* Allocate(int size) {
+    Datablock block(new char[size], size);
+    blocks_.push_back(block);
+    return block.data;
+  }
+
+  // We flush only at the end, because the writer wants
+  // random access to the blocks and once we hand the
+  // block over to the sink, we can't access it anymore.
+  // Also we don't write more than has been actually written
+  // to the blocks.
+  void Flush(size_t size) {
+    size_t size_written = 0;
+    size_t block_size;
+    for (int i = 0; i < blocks_.size(); ++i) {
+      block_size = min<size_t>(blocks_[i].size, size - size_written);
+      dest_->AppendAndTakeOwnership(blocks_[i].data, block_size,
+                                    &SnappySinkAllocator::Deleter, NULL);
+      size_written += block_size;
+    }
+    blocks_.clear();
+  }
+
+ private:
+  struct Datablock {
+    char* data;
+    size_t size;
+    Datablock(char* p, size_t s) : data(p), size(s) {}
+  };
+
+  static void Deleter(void* arg, const char* bytes, size_t size) {
+    delete[] bytes;
+  }
+
+  Sink* dest_;
+  std::vector<Datablock> blocks_;
+
+  // Note: copying this object is allowed
+};
+
+size_t UncompressAsMuchAsPossible(Source* compressed, Sink* uncompressed) {
+  SnappySinkAllocator allocator(uncompressed);
+  SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
+  InternalUncompress(compressed, &writer);
+  return writer.Produced();
+}
+
+bool Uncompress(Source* compressed, Sink* uncompressed) {
+  // Read the uncompressed length from the front of the compressed input
+  SnappyDecompressor decompressor(compressed);
+  uint32 uncompressed_len = 0;
+  if (!decompressor.ReadUncompressedLength(&uncompressed_len)) {
+    return false;
+  }
+
+  char c;
+  size_t allocated_size;
+  char* buf = uncompressed->GetAppendBufferVariable(
+      1, uncompressed_len, &c, 1, &allocated_size);
+
+  // If we can get a flat buffer, then use it, otherwise do block by block
+  // uncompression
+  if (allocated_size >= uncompressed_len) {
+    SnappyArrayWriter writer(buf);
+    bool result = InternalUncompressAllTags(
+        &decompressor, &writer, uncompressed_len);
+    uncompressed->Append(buf, writer.Produced());
+    return result;
+  } else {
+    SnappySinkAllocator allocator(uncompressed);
+    SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
+    return InternalUncompressAllTags(&decompressor, &writer, uncompressed_len);
+  }
+}
+
+} // end namespace snappy