Compiler Explorer

Source code

#include <string>
#include <cstdint>
#include <cstddef>
#include <cstring>
#include <array>
#include <cassert>
#include <arm_neon.h>

inline constexpr int16_t MakeEntry(int16_t len, int16_t offset) {
  return len - (offset << 8);
}

inline constexpr int16_t LengthMinusOffset(int data, int type) {
  return type == 3   ? 0xFF                    // copy-4 (or type == 3)
         : type == 2 ? MakeEntry(data + 1, 0)  // copy-2
         : type == 1 ? MakeEntry((data & 7) + 4, data >> 3)  // copy-1
         : data < 60 ? MakeEntry(data + 1, 1)  // note spurious offset.
                     : 0xFF;                   // long literal
}

inline constexpr int16_t LengthMinusOffset(uint8_t tag) {
  return LengthMinusOffset(tag >> 2, tag & 3);
}

template <size_t... Ints>
struct index_sequence {};

template <std::size_t N, size_t... Is>
struct make_index_sequence : make_index_sequence<N - 1, N - 1, Is...> {};

template <size_t... Is>
struct make_index_sequence<0, Is...> : index_sequence<Is...> {};

template <size_t... seq>
constexpr std::array<int16_t, 256> MakeTable(index_sequence<seq...>) {
  return std::array<int16_t, 256>{LengthMinusOffset(seq)...};
}

alignas(64) const std::array<int16_t, 256> kLengthMinusOffset =
    MakeTable(make_index_sequence<256>{});

inline uint32_t ExtractOffset(uint32_t val, size_t tag_type) {
  // For x86 non-static storage works better. For ARM static storage is better.
  // TODO(b/194925505): Once the array is recognized as a register, improve the
  // readability for x86.
#if 1 //defined(__x86_64__)
  constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
  uint16_t result;
  memcpy(&result,
         reinterpret_cast<const char*>(&kExtractMasksCombined) + 2 * tag_type,
         sizeof(result));
  return val & result;
#else
  static constexpr uint32_t kExtractMasks[4] = {0, 0xFF, 0xFFFF, 0};
  return val & kExtractMasks[tag_type];
#endif
};

constexpr int kSlopBytes = 64;
#define ZIPPY_PREDICT_FALSE(x) (__builtin_expect(x, 0))
#define ZIPPY_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
#define ZIPPY_ATTRIBUTE_ALWAYS_INLINE __attribute__((always_inline))

ZIPPY_ATTRIBUTE_ALWAYS_INLINE
size_t AdvanceToNextTag(const uint8_t** ip_p, size_t* tag) {
  const uint8_t*& ip = *ip_p;
  const size_t tag_type = *tag & 3;
  const bool is_literal = (tag_type == 0);
  if (is_literal) {
    size_t next_literal_tag = (*tag >> 2) + 1;
    *tag = ip[next_literal_tag];
    ip += next_literal_tag + 1;
  } else {
    *tag = ip[tag_type];
    ip += tag_type + 1;
  }
  return tag_type;
}

template <size_t... indexes>
inline constexpr std::array<char, sizeof...(indexes)> MakePatternMaskBytes(
    int index_offset, int pattern_size, index_sequence<indexes...>) {
  return {static_cast<char>((index_offset + indexes) % pattern_size)...};
}

using V128 = uint8x16_t;

inline V128 V128_Load(const V128* src) {
  return vld1q_u8(reinterpret_cast<const uint8_t*>(src));
}

inline V128 V128_LoadU(const V128* src) {
  return vld1q_u8(reinterpret_cast<const uint8_t*>(src));
}

inline void V128_StoreU(V128* dst, V128 val) {
  vst1q_u8(reinterpret_cast<uint8_t*>(dst), val);
}

inline V128 V128_Shuffle(V128 input, V128 shuffle_mask) {
  assert(vminvq_u8(shuffle_mask) >= 0 && vmaxvq_u8(shuffle_mask) <= 15);
  return vqtbl1q_u8(input, shuffle_mask);
}

inline V128 V128_DupChar(char c) { return vdupq_n_u8(c); }
// Computes the shuffle control mask bytes array for given pattern-sizes and
// returns an array.
template <size_t... pattern_sizes_minus_one>
inline constexpr std::array<std::array<char, sizeof(V128)>,
                            sizeof...(pattern_sizes_minus_one)>
MakePatternMaskBytesTable(int index_offset,
                          index_sequence<pattern_sizes_minus_one...>) {
  return {
      MakePatternMaskBytes(index_offset, pattern_sizes_minus_one + 1,
                           make_index_sequence</*indexes=*/sizeof(V128)>())...};
}

// This is an array of shuffle control masks that can be used as the source
// operand for PSHUFB to permute the contents of the destination XMM register
// into a repeating byte pattern.
alignas(16) constexpr std::array<std::array<char, sizeof(V128)>,
                                 16> pattern_generation_masks =
    MakePatternMaskBytesTable(
        /*index_offset=*/0,
        /*pattern_sizes_minus_one=*/make_index_sequence<16>());

// Similar to 'pattern_generation_masks', this table is used to "rotate" the
// pattern so that we can copy the *next 16 bytes* consistent with the pattern.
// Basically, pattern_reshuffle_masks is a continuation of
// pattern_generation_masks. It follows that, pattern_reshuffle_masks is same as
// pattern_generation_masks for offsets 1, 2, 4, 8 and 16.
alignas(16) constexpr std::array<std::array<char, sizeof(V128)>,
                                 16> pattern_reshuffle_masks =
    MakePatternMaskBytesTable(
        /*index_offset=*/16,
        /*pattern_sizes_minus_one=*/make_index_sequence<16>());

ZIPPY_ATTRIBUTE_ALWAYS_INLINE
static inline V128 LoadPattern(const char* src, const size_t pattern_size) {
  V128 generation_mask = V128_Load(reinterpret_cast<const V128*>(
      pattern_generation_masks[pattern_size - 1].data()));
  // Uninitialized bytes are masked out by the shuffle mask.
  return V128_Shuffle(V128_LoadU(reinterpret_cast<const V128*>(src)),
                      generation_mask);
}

ZIPPY_ATTRIBUTE_ALWAYS_INLINE
static inline std::pair<V128 /* pattern */, V128 /* reshuffle_mask */>
LoadPatternAndReshuffleMask(const char* src, const size_t pattern_size) {
  V128 pattern = LoadPattern(src, pattern_size);

// This mask will generate the next 16 bytes in-place. Doing so enables us to
  // write data by at most 4 V128_StoreU.
  //
  // For example, suppose pattern is:        abcdefabcdefabcd
  // Shuffling with this mask will generate: efabcdefabcdefab
  // Shuffling again will generate:          cdefabcdefabcdef
  V128 reshuffle_mask = V128_Load(reinterpret_cast<const V128*>(
      pattern_reshuffle_masks[pattern_size - 1].data()));
  return {pattern, reshuffle_mask};
}

inline bool Copy64BytesWithPatternExtension(char* dst, size_t offset) {
#if 1
  if (ZIPPY_PREDICT_TRUE(offset <= 16)) {
    switch (offset) {
      case 0:
        return false;
      case 1: {
        V128 pattern = V128_DupChar(dst[-1]);
        for (int i = 0; i < 4; i++) {
          V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern);
        }
        return true;
      }
      case 2:
      case 4:
      case 8:
      case 16: {
        V128 pattern = LoadPattern(dst - offset, offset);
        for (int i = 0; i < 4; i++) {
          V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern);
        }
        return true;
      }
      default: {
        auto pattern_and_reshuffle_mask =
            LoadPatternAndReshuffleMask(dst - offset, offset);
        V128 pattern = pattern_and_reshuffle_mask.first;
        V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
        for (int i = 0; i < 4; i++) {
          V128_StoreU(reinterpret_cast<V128*>(dst + 16 * i), pattern);
          pattern = V128_Shuffle(pattern, reshuffle_mask);
        }
        return true;
      }
    }
  }
#else
  if (ZIPPY_PREDICT_TRUE(offset < 16)) {
    if (ZIPPY_PREDICT_FALSE(offset == 0)) return false;
    // Extend the pattern to the first 16 bytes.
    for (int i = 0; i < 16; i++) dst[i] = dst[i - offset];
    // Find a multiple of pattern >= 16.
    static std::array<uint8_t, 16> pattern_sizes = []() {
      std::array<uint8_t, 16> res;
      for (int i = 1; i < 16; i++) res[i] = (16 / i + 1) * i;
      return res;
    }();
    offset = pattern_sizes[offset];
    for (int i = 1; i < 4; i++) {
      std::memcpy(dst + i * 16, dst + i * 16 - offset, 16);
    }
    return true;
  }
#endif  // ZIPPY_HAVE_VECTOR_BYTE_SHUFFLE

// Very rare.
  for (int i = 0; i < 4; i++) {
    std::memcpy(dst + i * 16, dst + i * 16 - offset, 16);
  }
  return true;
}

inline void MemCopy64(char* dst, const void* src, size_t size) {
  // Always copy this many bytes, test if we need to copy more.
  constexpr int kShortMemCopy = 32;
  // We're always allowed to copy 64 bytes, so if we exceed kShortMemCopy just
  // copy 64 rather than the exact amount.
  constexpr int kLongMemCopy = 64;

assert(size <= kLongMemCopy);
  assert(std::less_equal<const void*>()(static_cast<const char*>(src) + size,
                                        dst) ||
         std::less_equal<const void*>()(dst + size, src));

// We know that src and dst are at least size bytes apart. However, because we
  // might copy more than size bytes the copy still might overlap past size.
  // E.g. if src and dst appear consecutively in memory (src + size == dst).
  std::memmove(dst, src, kShortMemCopy);
  // Profiling shows that nearly all copies are short.
  if (ZIPPY_PREDICT_FALSE(size > kShortMemCopy)) {
    std::memmove(dst + kShortMemCopy,
                 static_cast<const uint8_t*>(src) + kShortMemCopy,
                 kLongMemCopy - kShortMemCopy);
  }
}

inline uint32_t UNALIGNED_LOAD32(const void *p) {
  uint32_t t;
  memcpy(&t, p, sizeof t);
  return t;
}

inline uint32_t UNALIGNED_LOAD16(const void *p) {
  uint16_t t;
  memcpy(&t, p, sizeof t);
  return t;
}

std::pair<const uint8_t*, ptrdiff_t> DecompressBranchless(
    const uint8_t* ip, const uint8_t* ip_limit, ptrdiff_t op, char* op_base,
    ptrdiff_t op_limit_min_slop) {
  // We unroll the inner loop twice so we need twice the spare room.
  op_limit_min_slop -= kSlopBytes;
  if (2 * (kSlopBytes + 1) < ip_limit - ip && op < op_limit_min_slop) {
    const uint8_t* const ip_limit_min_slop = ip_limit - 2 * kSlopBytes - 1;
    ip++;
    // ip points just past the tag and we are touching at maximum kSlopBytes
    // in an iteration.
    size_t tag = ip[-1];
    #if 0 //defined(__clang__) && defined(__aarch64__)
    // Workaround for https://bugs.llvm.org/show_bug.cgi?id=51317
    // when loading 1 byte, clang for aarch64 doesn't realize that it(ldrb)
    // comes with free zero-extension, so clang generates another
    // 'and xn, xm, 0xff' before it use that as the offset. This 'and' is
    // redundant and can be removed by adding this dummy asm, which gives
    // clang a hint that we're doing the zero-extension at the load.
    asm("" ::"r"(tag));
#endif
    do {
      // The throughput is limited by instructions, unrolling the inner loop
      // twice reduces the amount of instructions checking limits and also
      // leads to reduced mov's.
      for (int i = 0; i < 2; i++) {
        tag += 4;
        const uint8_t* old_ip = ip;
        assert(tag == ip[-1]);
        // For literals tag_type = 0, hence we will always obtain 0 from
        // ExtractLowBytes. For literals offset will thus be kLiteralOffset.
        ptrdiff_t len_min_offset = kLengthMinusOffset[tag - 4];
        size_t tag_type = AdvanceToNextTag(&ip, &tag);
        uint32_t next = UNALIGNED_LOAD32(old_ip);
        size_t len = len_min_offset & 0xFF;
        len_min_offset -= ExtractOffset(next, tag_type);
        if (ZIPPY_PREDICT_FALSE(len_min_offset > 0)) {
          if (ZIPPY_PREDICT_FALSE(len & 0x80)) {
            // Exceptional case (long literal or copy 4).
            // Actually doing the copy here is negatively impacting the main
            // loop due to compiler incorrectly allocating a register for
            // this fallback. Hence we just break.
          break_loop:
            ip = old_ip;
            goto exit;
          }
          // Only copy-1 or copy-2 tags can get here.
          assert(tag_type == 1 || tag_type == 2);
          std::ptrdiff_t delta = op + len_min_offset - len;
          // Guard against copies before the buffer start.
          if (ZIPPY_PREDICT_FALSE(delta < 0 ||
                                  !Copy64BytesWithPatternExtension(
                                      op_base + op, len - len_min_offset))) {
            goto break_loop;
          }
          op += len;
          continue;
        }
        std::ptrdiff_t delta = op + len_min_offset - len;
        if (ZIPPY_PREDICT_FALSE(delta < 0)) {
          // Due to the spurious offset in literals have this will trigger
          // at the start of a block when op is still smaller than 256.
          if (tag_type != 0) goto break_loop;
          MemCopy64(op_base + op, old_ip, len);
          op += len;
          continue;
        }

// For copies we need to copy from op_base + delta, for literals
        // we need to copy from ip instead of from the stream.
        const void* from =
            tag_type ? reinterpret_cast<void*>(op_base + delta) : old_ip;
        MemCopy64(op_base + op, from, len);
        op += len;
      }
    } while (ip < ip_limit_min_slop && op < op_limit_min_slop);
  exit:
    ip--;
    assert(ip <= ip_limit);
  }
  return {ip, op};
}