Fix benchmark C++11 code

PiperOrigin-RevId: 913030038

.github/workflows/riscv64-qemu-test.yaml

@@ -0,0 +1,43 @@
+name: riscv64-qemu-test
+
+on: [push, pull_request]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    env: 
+      RISCV_CROSSCOMPILE: "ON"
+      riscv_gnu_toolchain_download_path: https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2025.07.03/riscv64-glibc-ubuntu-24.04-gcc-nightly-2025.07.03-nightly.tar.xz
+      RISCV_PATH: /opt/riscv
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        submodules: recursive
+    
+    - name: Install dependencies
+      run: |
+        sudo apt update
+        sudo apt install -y --no-install-recommends \
+          qemu-user qemu-user-static \
+          build-essential \
+          cmake \
+          git
+        sudo mkdir -p $RISCV_PATH
+        wget ${riscv_gnu_toolchain_download_path} -O riscv-toolchain.tar.xz
+        sudo tar -xvf riscv-toolchain.tar.xz -C $RISCV_PATH --strip-components=1
+        sudo sed -i "s|libdir='/mnt/riscv/riscv64-unknown-linux-gnu/lib'|libdir='$RISCV_PATH/riscv64-unknown-linux-gnu/lib'|g" $RISCV_PATH/riscv64-unknown-linux-gnu/lib/libatomic.la
+
+    - name: Build and Run Unit Tests
+      run: |
+        export PATH=$RISCV_PATH/bin:$PATH
+        export LD_LIBRARY_PATH="/opt/riscv/lib:$LD_LIBRARY_PATH"
+        export QEMU_LD_PREFIX=$RISCV_PATH/sysroot
+        mkdir build && cd build
+        cmake -DCMAKE_BUILD_TYPE=Release ../
+        make -j$(nproc)
+        make test
+
+    - name: Run Benchmark
+      run: ./build/snappy_benchmark
+      working-directory: ./

BUILD.bazel

@@ -26,10 +26,20 @@
 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+load("@bazel_skylib//rules:expand_template.bzl", "expand_template")
+load("@bazel_skylib//rules:write_file.bzl", "write_file")
+load("@rules_cc//cc:cc_library.bzl", "cc_library")
+load("@rules_cc//cc:cc_test.bzl", "cc_test")
+
 package(default_visibility = ["//visibility:public"])
 
 licenses(["notice"])
 
+exports_files([
+    "COPYING",
+    "COPYING.notestdata",
+])
+
 SNAPPY_VERSION = (1, 2, 2)
 
 config_setting(
@@ -52,8 +62,10 @@ cc_library(
     name = "snappy-stubs-internal",
     srcs = ["snappy-stubs-internal.cc"],
     hdrs = ["snappy-stubs-internal.h"],
-    deps = [
+    implementation_deps = [
         ":config",
+    ],
+    deps = [
         ":snappy-stubs-public",
     ],
 )
@@ -75,8 +87,10 @@ cc_library(
             "-Wno-sign-compare",
         ],
     }),
-    deps = [
+    implementation_deps = [
         ":config",
+    ],
+    deps = [
         ":snappy-stubs-internal",
         ":snappy-stubs-public",
     ],
@@ -133,10 +147,10 @@ cc_test(
 )
 
 # Generate a config.h similar to what cmake would produce.
-genrule(
+write_file(
     name = "config_h",
-    outs = ["config.h"],
+    out = "config.h",
-    cmd = """cat <<EOF >$@
+    content = """\
 #define HAVE_STDDEF_H 1
 #define HAVE_STDINT_H 1
 #ifdef __has_builtin
@@ -193,19 +207,17 @@ genrule(
 #    define SNAPPY_IS_BIG_ENDIAN 1
 #  endif
 #endif
-EOF
+""".splitlines(),
-""",
 )
 
-genrule(
+expand_template(
     name = "snappy_stubs_public_h",
-    srcs = ["snappy-stubs-public.h.in"],
+    out = "snappy-stubs-public.h",
-    outs = ["snappy-stubs-public.h"],
+    substitutions = {
-    # Assume sys/uio.h is available on non-Windows.
+        "${HAVE_SYS_UIO_H_01}": "!_WIN32",
-    # Set the version numbers.
+        "${PROJECT_VERSION_MAJOR}": str(SNAPPY_VERSION[0]),
-    cmd = ("""sed -e 's/$${HAVE_SYS_UIO_H_01}/!_WIN32/g' \
+        "${PROJECT_VERSION_MINOR}": str(SNAPPY_VERSION[1]),
-           -e 's/$${PROJECT_VERSION_MAJOR}/%d/g' \
+        "${PROJECT_VERSION_PATCH}": str(SNAPPY_VERSION[2]),
-           -e 's/$${PROJECT_VERSION_MINOR}/%d/g' \
+    },
-           -e 's/$${PROJECT_VERSION_PATCH}/%d/g' \
+    template = "snappy-stubs-public.h.in",
-    $< >$@""" % SNAPPY_VERSION),
 )

CMakeLists.txt

@@ -155,10 +155,20 @@ int main() {
   return __builtin_expect(0, 1);
 }" HAVE_BUILTIN_EXPECT)
 
+# Check if the built-in __builtin_ctz (count trailing zeros) is available.
+# Require either a non-RISC-V target, or a RISC-V core that implements
+# the Zbb bit-manipulation extension where ctz is guaranteed.
 check_cxx_source_compiles("
-int main() {
+#ifdef __riscv
-  return __builtin_ctzll(0);
+  #ifdef __riscv_zbb
-}" HAVE_BUILTIN_CTZ)
+    int main() { return __builtin_ctzll(0); }
+  #else
+    #error \"ZBB not enabled in current config\"
+  #endif
+#else
+  int main() { return __builtin_ctzll(0); }
+#endif
+" HAVE_BUILTIN_CTZ)
 
 check_cxx_source_compiles("
 int main() {
@@ -216,6 +226,31 @@ int main() {
   return 0;
 }" SNAPPY_HAVE_NEON)
 
+#check RVV 1.0 need __riscv_ prefix
+check_cxx_source_compiles("
+  #include <riscv_vector.h>
+  #include <stdint.h>
+  #include <stddef.h>
+  int main() {
+    uint8_t val = 3, dup[8];
+    size_t vl = __riscv_vsetvl_e8m1(8);
+    vuint8m1_t v = __riscv_vmv_v_x_u8m1(val, vl);
+    return 0;
+  }" SNAPPY_RVV_1)
+
+
+#check RVV 0.7.1 not  __riscv_ prefix
+check_cxx_source_compiles("
+  #include <riscv_vector.h>
+  #include <stdint.h>
+  #include <stddef.h>
+  int main() {
+    uint8_t val = 3, dup[8];
+    size_t vl = vsetvl_e8m1(8);
+    vuint8m1_t v = vmv_v_x_u8m1(val, vl);
+    return 0;
+  }" SNAPPY_RVV_0_7)
+
 include(CheckSymbolExists)
 check_symbol_exists("mmap" "sys/mman.h" HAVE_FUNC_MMAP)
 check_symbol_exists("sysconf" "unistd.h" HAVE_FUNC_SYSCONF)

COPYING.notestdata

@@ -0,0 +1,28 @@
+Copyright 2011, Google Inc.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

MODULE.bazel

@@ -4,20 +4,23 @@ module(
     compatibility_level = 1,
 )
 
+bazel_dep(name = "rules_cc", version = "0.2.4")
+bazel_dep(name = "bazel_skylib", version = "1.8.2")
+
 bazel_dep(
     name = "googletest",
-    version = "1.14.0.bcr.1",
+    version = "1.17.0.bcr.2",
     dev_dependency = True,
     repo_name = "com_google_googletest",
 )
 bazel_dep(
     name = "google_benchmark",
-    version = "1.9.0",
+    version = "1.9.5",
     dev_dependency = True,
     repo_name = "com_google_benchmark",
 )
 
 bazel_dep(
     name = "platforms",
-    version = "0.0.9",
+    version = "1.0.0",
 )

cmake/config.h.in

@@ -58,6 +58,12 @@
 /* Define to 1 if you target processors with NEON and have <arm_neon.h>. */
 #cmakedefine01 SNAPPY_HAVE_NEON
 
+/* Define to 1 if you target processors with RVV1.0 and have <riscv_vector.h>. */
+#cmakedefine01 SNAPPY_RVV_1
+
+/* Define to 1 if you target processors with RVV0.7 and have <riscv_vector.h>. */
+#cmakedefine01 SNAPPY_RVV_0_7
+
 /* Define to 1 if you have <arm_neon.h> and <arm_acle.h> and want to optimize
    compression speed by using __crc32cw from <arm_acle.h>. */
 #cmakedefine01 SNAPPY_HAVE_NEON_CRC32

snappy-internal.h

@@ -46,6 +46,23 @@
 #include <arm_neon.h>
 #endif
 
+#if SNAPPY_RVV_1 || SNAPPY_RVV_0_7
+#define SNAPPY_HAVE_RVV 1
+#include <riscv_vector.h>
+#else
+#define SNAPPY_HAVE_RVV 0
+#endif
+
+#ifdef SNAPPY_RVV_1
+#define VSETVL_E8M2 __riscv_vsetvl_e8m2
+#define VLE8_V_U8M2 __riscv_vle8_v_u8m2
+#define VSE8_V_U8M2 __riscv_vse8_v_u8m2
+#elif SNAPPY_RVV_0_7
+#define VSETVL_E8M2 vsetvl_e8m2
+#define VLE8_V_U8M2 vle8_v_u8m2
+#define VSE8_V_U8M2 vse8_v_u8m2
+#endif
+
 #if SNAPPY_HAVE_SSSE3 || SNAPPY_HAVE_NEON
 #define SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE 1
 #else
@@ -110,6 +127,8 @@ inline V128 V128_Shuffle(V128 input, V128 shuffle_mask) {
 }
 
 inline V128 V128_DupChar(char c) { return vdupq_n_u8(c); }
+
+
 #endif
 #endif  // SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
 
@@ -172,9 +191,10 @@ char* CompressFragment(const char* input,
 // loading from s2 + n.
 //
 // Separate implementation for 64-bit, little-endian cpus.
+// riscv and little-endian cpu choose this routinue can be done faster too.
 #if !SNAPPY_IS_BIG_ENDIAN && \
     (defined(__x86_64__) || defined(_M_X64) || defined(ARCH_PPC) || \
-     defined(ARCH_ARM))
+     defined(__aarch64__) || (defined(__riscv) && (__riscv_xlen == 64)))
 static inline std::pair<size_t, bool> FindMatchLength(const char* s1,
                                                       const char* s2,
                                                       const char* s2_limit,

snappy-stubs-internal.h

@@ -304,6 +304,7 @@ class Bits {
   void operator=(const Bits&);
 };
 
+// In RISC-V,  CLZ  is supported by instructions from the ZBB bit-manipulation extension.
 #if HAVE_BUILTIN_CTZ
 
 inline int Bits::Log2FloorNonZero(uint32_t n) {
@@ -393,6 +394,7 @@ inline int Bits::FindLSBSetNonZero(uint32_t n) {
 
 #endif  // End portable versions.
 
+// In RISC-V,  CLZ  is supported by instructions from the ZBB bit-manipulation extension.
 #if HAVE_BUILTIN_CTZ
 
 inline int Bits::FindLSBSetNonZero64(uint64_t n) {

snappy.cc

@@ -74,7 +74,6 @@
 #include <cstdint>
 #include <cstdio>
 #include <cstring>
-#include <functional>
 #include <memory>
 #include <string>
 #include <utility>
@@ -281,6 +280,8 @@ inline char* IncrementalCopySlow(const char* src, char* op,
 // 4, 5, 0, 1, 2, 3, 4, 5, 0, 1}. These byte index sequences are generated by
 // calling MakePatternMaskBytes(0, 6, index_sequence<16>()) and
 // MakePatternMaskBytes(16, 6, index_sequence<16>()) respectively.
+
+
 template <size_t... indexes>
 inline constexpr std::array<char, sizeof...(indexes)> MakePatternMaskBytes(
     int index_offset, int pattern_size, index_sequence<indexes...>) {
@@ -298,7 +299,6 @@ MakePatternMaskBytesTable(int index_offset,
       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.
@@ -494,7 +494,6 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
           LoadPatternAndReshuffleMask(src, pattern_size);
       V128 pattern = pattern_and_reshuffle_mask.first;
       V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
-
       // There is at least one, and at most four 16-byte blocks. Writing four
       // conditionals instead of a loop allows FDO to layout the code with
       // respect to the actual probabilities of each length.
@@ -521,7 +520,6 @@ inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
           LoadPatternAndReshuffleMask(src, pattern_size);
       V128 pattern = pattern_and_reshuffle_mask.first;
       V128 reshuffle_mask = pattern_and_reshuffle_mask.second;
-
       // This code path is relatively cold however so we save code size
       // by avoiding unrolling and vectorizing.
       //
@@ -1227,7 +1225,7 @@ inline bool Copy64BytesWithPatternExtension(ptrdiff_t dst, size_t offset) {
 void MemCopy64(char* dst, const void* src, size_t size) {
   // Always copy this many bytes.  If that's below size then copy the full 64.
   constexpr int kShortMemCopy = 32;
-
+  (void)kShortMemCopy;
   assert(size <= 64);
   assert(std::less_equal<const void*>()(static_cast<const char*>(src) + size,
                                         dst) ||
@@ -1246,6 +1244,27 @@ void MemCopy64(char* dst, const void* src, size_t size) {
     data = _mm256_lddqu_si256(static_cast<const __m256i *>(src) + 1);
     _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst) + 1, data);
   }
+  // RVV acceleration available on RISC-V when compiled with -march=rv64gcv
+#elif defined(__riscv) && SNAPPY_HAVE_RVV
+  // Cast pointers to the type we will operate on.
+  unsigned char* dst_ptr = reinterpret_cast<unsigned char*>(dst);
+  const unsigned char* src_ptr = reinterpret_cast<const unsigned char*>(src);
+  size_t remaining_bytes = size;
+  // Loop as long as there are bytes remaining to be copied.
+  while (remaining_bytes > 0) {
+    // Set vector configuration: e8 (8-bit elements), m2 (LMUL=2).
+    // Use e8m2 configuration to maximize throughput.
+    size_t vl = VSETVL_E8M2(remaining_bytes);
+    // Load data from the current source pointer.
+    vuint8m2_t vec = VLE8_V_U8M2(src_ptr, vl);
+    // Store data to the current destination pointer.
+    VSE8_V_U8M2(dst_ptr, vec, vl);
+    // Update pointers and the remaining count.
+    src_ptr += vl;
+    dst_ptr += vl;
+    remaining_bytes -= vl;
+  }
+
 #else
   std::memmove(dst, src, kShortMemCopy);
   // Profiling shows that nearly all copies are short.
@@ -1345,19 +1364,51 @@ inline size_t AdvanceToNextTagX86Optimized(const uint8_t** ip_p, size_t* tag) {
   return tag_type;
 }
 
+SNAPPY_ATTRIBUTE_ALWAYS_INLINE
+inline size_t AdvanceToNextTagRVOptimized(const uint8_t** ip_p, size_t* tag) {
+  const uint8_t*& ip = *ip_p;
+  // This section is crucial for the throughput of the decompression loop.
+  // The latency of an iteration is fundamentally constrained by the data chain on ip:
+  // ip -> c = *tag -> literal_len = c >> 2, tag_type = c & 3
+  //      -> literal_advance = literal_len + 2, copy_advance = tag_type + 1
+  //      -> next_ip = ip + literal_advance  OR  ip + copy_advance (literal vs copy)
+  //      -> *tag = byte at (next_ip - 1); ip = next_ip
+  //
+  // Base RISC-V has no x86-style cmov and no AArch64 csinc on the same shape; this
+  // computes both candidate advances and both load offsets, then selects with
+  // (is_literal ? ... : ...). With the Zicond extension (czero.eqz / czero.nez), those
+  // selections typically lower to branchless conditional-zero ops instead of a
+  // hard-to-predict literal/copy branch, which is why this form tends to win there.
+  const size_t literal_len = *tag >> 2;
+  const size_t tag_type = *tag & 3;
+  const bool is_literal = (tag_type == 0);
+  const size_t copy_advance = tag_type + 1;
+  const size_t literal_advance = literal_len + 2;
+  const uint8_t* next_ip = is_literal ? (ip + literal_advance) : (ip + copy_advance);
+  *tag = is_literal ? ip[literal_advance - 1] : ip[copy_advance - 1];
+  ip = next_ip;
+  return tag_type;
+}
+
 // Extract the offset for copy-1 and copy-2 returns 0 for literals or copy-4.
 inline uint32_t ExtractOffset(uint32_t val, size_t tag_type) {
-  // For x86 non-static storage works better. For ARM static storage is better.
+  // For Arm non-static storage works better. For x86 static storage is better.
   // TODO: Once the array is recognized as a register, improve the
   // readability for x86.
 #if defined(__x86_64__)
-  constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
+  static constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
   uint16_t result;
   memcpy(&result,
          reinterpret_cast<const char*>(&kExtractMasksCombined) + 2 * tag_type,
          sizeof(result));
   return val & result;
-#elif defined(__aarch64__)
+  // For AArch64 and RISC-V, use a bit-twiddling trick to extract the mask from a
+  // single combined constant instead of a lookup table. The constant packs multiple
+  // 16-bit masks based on tag_type (see implementation below). The code calculates
+  // the shift amount from tag_type, right-shifts the constant to move the desired
+  // mask to the LSB position, then extracts it with & 0xFFFF. This branchless
+  // approach is often more performant on modern CPUs.
+#elif defined(__aarch64__) || (defined(__riscv) && (__riscv_xlen == 64))
   constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
   return val & static_cast<uint32_t>(
       (kExtractMasksCombined >> (tag_type * 16)) & 0xFFFF);
@@ -1421,6 +1472,11 @@ std::pair<const uint8_t*, ptrdiff_t> DecompressBranchless(
         // We never need more than 16 bits. Doing a Load16 allows the compiler
         // to elide the masking operation in ExtractOffset.
         next = LittleEndian::Load16(old_ip);
+#elif defined(__riscv)
+        size_t tag_type = AdvanceToNextTagRVOptimized(&ip, &tag);
+        // We never need more than 16 bits. Doing a Load16 allows the compiler
+        // to elide the masking operation in ExtractOffset.
+        next = LittleEndian::Load16(old_ip);
 #else
         size_t tag_type = AdvanceToNextTagX86Optimized(&ip, &tag);
         next = LittleEndian::Load32(old_ip);
@@ -1779,13 +1835,13 @@ static bool InternalUncompressAllTags(SnappyDecompressor* decompressor,
                                       Writer* writer, uint32_t compressed_len,
                                       uint32_t uncompressed_len) {
     int token = 0;
-  Report(token, "snappy_uncompress", compressed_len, uncompressed_len);
 
   writer->SetExpectedLength(uncompressed_len);
 
   // Process the entire input
   decompressor->DecompressAllTags(writer);
   writer->Flush();
+  Report(token, "snappy_uncompress", compressed_len, uncompressed_len);
   return (decompressor->eof() && writer->CheckLength());
 }
 
@@ -1800,11 +1856,9 @@ size_t Compress(Source* reader, Sink* writer) {
 
 size_t Compress(Source* reader, Sink* writer, CompressionOptions options) {
   assert(options.level == 1 || options.level == 2);
-  int token = 0;
   size_t written = 0;
   size_t N = reader->Available();
   assert(N <= 0xFFFFFFFFu);
-  const size_t uncompressed_size = N;
   char ulength[Varint::kMax32];
   char* p = Varint::Encode32(ulength, N);
   writer->Append(ulength, p - ulength);
@@ -1865,14 +1919,13 @@ size_t Compress(Source* reader, Sink* writer, CompressionOptions options) {
             fragment, fragment_size, dest, table, table_size >> 1,
             table + (table_size >> 1), table_size >> 1);
       }
+
     writer->Append(dest, end - dest);
     written += (end - dest);
 
     N -= num_to_read;
     reader->Skip(pending_advance);
   }
-
-  Report(token, "snappy_compress", written, uncompressed_size);
   return written;
 }
 

snappy.h

@@ -67,6 +67,7 @@ namespace snappy {
       constexpr CompressionOptions() = default;
       constexpr CompressionOptions(int compression_level)
           : level(compression_level) {}
+
     static constexpr int MinCompressionLevel() { return 1; }
     static constexpr int MaxCompressionLevel() { return 2; }
     static constexpr int DefaultCompressionLevel() { return 1; }
@@ -158,6 +159,7 @@ namespace snappy {
   //    RawCompress(input, input_length, output, &output_length);
   //    ... Process(output, output_length) ...
   //    delete [] output;
+  // First version is to preserve ABI.
   void RawCompress(const char* input, size_t input_length, char* compressed,
                    size_t* compressed_length);
   void RawCompress(const char* input, size_t input_length, char* compressed,
@@ -166,6 +168,7 @@ namespace snappy {
   // Same as `RawCompress` above but taking an `iovec` array as input. Note that
   // `uncompressed_length` is the total number of bytes to be read from the
   // elements of `iov` (_not_ the number of elements in `iov`).
+  // First version is to preserve ABI.
   void RawCompressFromIOVec(const struct iovec* iov, size_t uncompressed_length,
                             char* compressed, size_t* compressed_length);
   void RawCompressFromIOVec(const struct iovec* iov, size_t uncompressed_length,