grpc 1.61.0 → 1.62.0

data/third_party/abseil-cpp/absl/crc/internal/cpu_detect.cc

@@ -189,8 +189,14 @@ CpuType GetAmdCpuType() {
       break;
     case 0x19:
       switch (model_num) {
+        case 0x0:  // Stepping Ax
         case 0x1:  // Stepping B0
           return CpuType::kAmdMilan;
+        case 0x10:  // Stepping A0
+        case 0x11:  // Stepping B0
+          return CpuType::kAmdGenoa;
+        case 0x44:  // Stepping A0
+          return CpuType::kAmdRyzenV3000;
         default:
           return CpuType::kUnknown;
       }
@@ -237,8 +243,26 @@ CpuType GetCpuType() {
     ABSL_INTERNAL_AARCH64_ID_REG_READ(MIDR_EL1, midr);
     uint32_t implementer = (midr >> 24) & 0xff;
     uint32_t part_number = (midr >> 4) & 0xfff;
-    if (implementer == 0x41 && part_number == 0xd0c) {
-      return CpuType::kArmNeoverseN1;
+    switch (implementer) {
+      case 0x41:
+        switch (part_number) {
+          case 0xd0c: return CpuType::kArmNeoverseN1;
+          case 0xd40: return CpuType::kArmNeoverseV1;
+          case 0xd49: return CpuType::kArmNeoverseN2;
+          case 0xd4f: return CpuType::kArmNeoverseV2;
+          default:
+            return CpuType::kUnknown;
+        }
+        break;
+      case 0xc0:
+        switch (part_number) {
+          case 0xac3: return CpuType::kAmpereSiryn;
+          default:
+            return CpuType::kUnknown;
+        }
+        break;
+      default:
+        return CpuType::kUnknown;
     }
   }
   return CpuType::kUnknown;

data/third_party/abseil-cpp/absl/crc/internal/cpu_detect.h

@@ -29,6 +29,8 @@ enum class CpuType {
   kAmdRome,
   kAmdNaples,
   kAmdMilan,
+  kAmdGenoa,
+  kAmdRyzenV3000,
   kIntelCascadelakeXeon,
   kIntelSkylakeXeon,
   kIntelBroadwell,
@@ -37,6 +39,10 @@ enum class CpuType {
   kIntelSandybridge,
   kIntelWestmere,
   kArmNeoverseN1,
+  kArmNeoverseV1,
+  kAmpereSiryn,
+  kArmNeoverseN2,
+  kArmNeoverseV2
 };
 
 // Returns the type of host CPU this code is running on.  Returns kUnknown if

data/third_party/abseil-cpp/absl/crc/internal/crc32_x86_arm_combined_simd.h

@@ -59,6 +59,8 @@ namespace crc_internal {
 #if defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
 using V128 = uint64x2_t;
 #else
+// Note: Do not use __m128i_u, it is not portable.
+// Use V128_LoadU() perform an unaligned load from __m128i*.
 using V128 = __m128i;
 #endif
 
@@ -78,6 +80,9 @@ V128 V128_Load(const V128* src);
 // Load 128 bits of integer data. |src| does not need to be aligned.
 V128 V128_LoadU(const V128* src);
 
+// Store 128 bits of integer data. |src| must be 16-byte aligned.
+void V128_Store(V128* dst, V128 data);
+
 // Polynomially multiplies the high 64 bits of |l| and |r|.
 V128 V128_PMulHi(const V128 l, const V128 r);
 
@@ -109,6 +114,10 @@ V128 V128_ShiftRight(const V128 l);
 template <int imm>
 int V128_Extract32(const V128 l);
 
+// Extracts a 64-bit integer from |l|, selected with |imm|.
+template <int imm>
+uint64_t V128_Extract64(const V128 l);
+
 // Extracts the low 64 bits from V128.
 int64_t V128_Low64(const V128 l);
 
@@ -139,6 +148,8 @@ inline V128 V128_Load(const V128* src) { return _mm_load_si128(src); }
 
 inline V128 V128_LoadU(const V128* src) { return _mm_loadu_si128(src); }
 
+inline void V128_Store(V128* dst, V128 data) { _mm_store_si128(dst, data); }
+
 inline V128 V128_PMulHi(const V128 l, const V128 r) {
   return _mm_clmulepi64_si128(l, r, 0x11);
 }
@@ -173,6 +184,11 @@ inline int V128_Extract32(const V128 l) {
   return _mm_extract_epi32(l, imm);
 }
 
+template <int imm>
+inline uint64_t V128_Extract64(const V128 l) {
+  return static_cast<uint64_t>(_mm_extract_epi64(l, imm));
+}
+
 inline int64_t V128_Low64(const V128 l) { return _mm_cvtsi128_si64(l); }
 
 inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
@@ -203,10 +219,14 @@ inline V128 V128_LoadU(const V128* src) {
   return vld1q_u64(reinterpret_cast<const uint64_t*>(src));
 }
 
+inline void V128_Store(V128* dst, V128 data) {
+  vst1q_u64(reinterpret_cast<uint64_t*>(dst), data);
+}
+
 // Using inline assembly as clang does not generate the pmull2 instruction and
 // performance drops by 15-20%.
-// TODO(b/193678732): Investigate why the compiler decides not to generate
-// such instructions and why it becomes so much worse.
+// TODO(b/193678732): Investigate why there is a slight performance hit when
+// using intrinsics instead of inline assembly.
 inline V128 V128_PMulHi(const V128 l, const V128 r) {
   uint64x2_t res;
   __asm__ __volatile__("pmull2 %0.1q, %1.2d, %2.2d \n\t"
@@ -215,10 +235,14 @@ inline V128 V128_PMulHi(const V128 l, const V128 r) {
   return res;
 }
 
+// TODO(b/193678732): Investigate why the compiler decides to move the constant
+// loop multiplicands from GPR to Neon registers every loop iteration.
 inline V128 V128_PMulLow(const V128 l, const V128 r) {
-  return reinterpret_cast<V128>(vmull_p64(
-      reinterpret_cast<poly64_t>(vget_low_p64(vreinterpretq_p64_u64(l))),
-      reinterpret_cast<poly64_t>(vget_low_p64(vreinterpretq_p64_u64(r)))));
+  uint64x2_t res;
+  __asm__ __volatile__("pmull %0.1q, %1.1d, %2.1d \n\t"
+                       : "=w"(res)
+                       : "w"(l), "w"(r));
+  return res;
 }
 
 inline V128 V128_PMul01(const V128 l, const V128 r) {
@@ -252,6 +276,11 @@ inline int V128_Extract32(const V128 l) {
   return vgetq_lane_s32(vreinterpretq_s32_u64(l), imm);
 }
 
+template <int imm>
+inline uint64_t V128_Extract64(const V128 l) {
+  return vgetq_lane_u64(l, imm);
+}
+
 inline int64_t V128_Low64(const V128 l) {
   return vgetq_lane_s64(vreinterpretq_s64_u64(l), 0);
 }

data/third_party/abseil-cpp/absl/crc/internal/crc_memcpy.h

@@ -20,12 +20,15 @@
 
 #include "absl/base/config.h"
 #include "absl/crc/crc32c.h"
+#include "absl/crc/internal/crc32_x86_arm_combined_simd.h"
 
 // Defined if the class AcceleratedCrcMemcpyEngine exists.
-#if defined(__x86_64__) && defined(__SSE4_2__)
-#define ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE 1
-#elif defined(_MSC_VER) && defined(__AVX__)
+// TODO(b/299127771): Consider relaxing the pclmul requirement once the other
+// intrinsics are conditionally compiled without it.
+#if defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
 #define ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE 1
+#elif defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
+#define ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE 1
 #endif
 
 namespace absl {

data/third_party/abseil-cpp/absl/crc/internal/crc_memcpy_fallback.cc

@@ -54,7 +54,8 @@ absl::crc32c_t FallbackCrcMemcpyEngine::Compute(void* __restrict dst,
 }
 
 // Compile the following only if we don't have
-#ifndef ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE
+#if !defined(ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE) && \
+    !defined(ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE)
 
 CrcMemcpy::ArchSpecificEngines CrcMemcpy::GetArchSpecificEngines() {
   CrcMemcpy::ArchSpecificEngines engines;
@@ -68,7 +69,8 @@ std::unique_ptr<CrcMemcpyEngine> CrcMemcpy::GetTestEngine(int /*vector*/,
   return std::make_unique<FallbackCrcMemcpyEngine>();
 }
 
-#endif  // ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE
+#endif  // !ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE &&
+        // !ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE
 
 }  // namespace crc_internal
 ABSL_NAMESPACE_END

data/third_party/abseil-cpp/absl/crc/internal/{crc_memcpy_x86_64.cc → crc_memcpy_x86_arm_combined.cc}

@@ -12,9 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// Simultaneous memcopy and CRC-32C for x86-64.  Uses integer registers because
-// XMM registers do not support the CRC instruction (yet).  While copying,
-// compute the running CRC of the data being copied.
+// Simultaneous memcopy and CRC-32C for x86-64 and ARM 64. Uses integer
+// registers because XMM registers do not support the CRC instruction (yet).
+// While copying, compute the running CRC of the data being copied.
 //
 // It is assumed that any CPU running this code has SSE4.2 instructions
 // available (for CRC32C).  This file will do nothing if that is not true.
@@ -49,17 +49,20 @@
 #include <array>
 #include <cstddef>
 #include <cstdint>
-#include <type_traits>
+#include <cstring>
+#include <memory>
 
-#include "absl/base/dynamic_annotations.h"
+#include "absl/base/config.h"
 #include "absl/base/optimization.h"
 #include "absl/base/prefetch.h"
 #include "absl/crc/crc32c.h"
 #include "absl/crc/internal/cpu_detect.h"
+#include "absl/crc/internal/crc32_x86_arm_combined_simd.h"
 #include "absl/crc/internal/crc_memcpy.h"
 #include "absl/strings/string_view.h"
 
-#ifdef ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE
+#if defined(ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE) || \
+    defined(ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE)
 
 namespace absl {
 ABSL_NAMESPACE_BEGIN
@@ -74,7 +77,7 @@ inline crc32c_t ShortCrcCopy(char* dst, const char* src, std::size_t length,
   uint32_t crc_uint32 = static_cast<uint32_t>(crc);
   for (std::size_t i = 0; i < length; i++) {
     uint8_t data = *reinterpret_cast<const uint8_t*>(src);
-    crc_uint32 = _mm_crc32_u8(crc_uint32, data);
+    crc_uint32 = CRC32_u8(crc_uint32, data);
     *reinterpret_cast<uint8_t*>(dst) = data;
     ++src;
     ++dst;
@@ -82,36 +85,35 @@ inline crc32c_t ShortCrcCopy(char* dst, const char* src, std::size_t length,
   return crc32c_t{crc_uint32};
 }
 
-constexpr size_t kIntLoadsPerVec = sizeof(__m128i) / sizeof(uint64_t);
+constexpr size_t kIntLoadsPerVec = sizeof(V128) / sizeof(uint64_t);
 
 // Common function for copying the tails of multiple large regions.
 template <size_t vec_regions, size_t int_regions>
 inline void LargeTailCopy(crc32c_t* crcs, char** dst, const char** src,
                           size_t region_size, size_t copy_rounds) {
-  std::array<__m128i, vec_regions> data;
+  std::array<V128, vec_regions> data;
   std::array<uint64_t, kIntLoadsPerVec * int_regions> int_data;
 
   while (copy_rounds > 0) {
     for (size_t i = 0; i < vec_regions; i++) {
       size_t region = i;
 
-      auto* vsrc =
-          reinterpret_cast<const __m128i*>(*src + region_size * region);
-      auto* vdst = reinterpret_cast<__m128i*>(*dst + region_size * region);
+      auto* vsrc = reinterpret_cast<const V128*>(*src + region_size * region);
+      auto* vdst = reinterpret_cast<V128*>(*dst + region_size * region);
 
       // Load the blocks, unaligned
-      data[i] = _mm_loadu_si128(vsrc);
+      data[i] = V128_LoadU(vsrc);
 
       // Store the blocks, aligned
-      _mm_store_si128(vdst, data[i]);
+      V128_Store(vdst, data[i]);
 
       // Compute the running CRC
       crcs[region] = crc32c_t{static_cast<uint32_t>(
-          _mm_crc32_u64(static_cast<uint32_t>(crcs[region]),
-                        static_cast<uint64_t>(_mm_extract_epi64(data[i], 0))))};
+          CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                    static_cast<uint64_t>(V128_Extract64<0>(data[i]))))};
       crcs[region] = crc32c_t{static_cast<uint32_t>(
-          _mm_crc32_u64(static_cast<uint32_t>(crcs[region]),
-                        static_cast<uint64_t>(_mm_extract_epi64(data[i], 1))))};
+          CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                    static_cast<uint64_t>(V128_Extract64<1>(data[i]))))};
     }
 
     for (size_t i = 0; i < int_regions; i++) {
@@ -125,7 +127,7 @@ inline void LargeTailCopy(crc32c_t* crcs, char** dst, const char** src,
         size_t data_index = i * kIntLoadsPerVec + j;
 
         int_data[data_index] = *(usrc + j);
-        crcs[region] = crc32c_t{static_cast<uint32_t>(_mm_crc32_u64(
+        crcs[region] = crc32c_t{static_cast<uint32_t>(CRC32_u64(
             static_cast<uint32_t>(crcs[region]), int_data[data_index]))};
 
         *(udst + j) = int_data[data_index];
@@ -133,8 +135,8 @@ inline void LargeTailCopy(crc32c_t* crcs, char** dst, const char** src,
     }
 
     // Increment pointers
-    *src += sizeof(__m128i);
-    *dst += sizeof(__m128i);
+    *src += sizeof(V128);
+    *dst += sizeof(V128);
     --copy_rounds;
   }
 }
@@ -158,8 +160,9 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
     void* __restrict dst, const void* __restrict src, std::size_t length,
     crc32c_t initial_crc) const {
   constexpr std::size_t kRegions = vec_regions + int_regions;
+  static_assert(kRegions > 0, "Must specify at least one region.");
   constexpr uint32_t kCrcDataXor = uint32_t{0xffffffff};
-  constexpr std::size_t kBlockSize = sizeof(__m128i);
+  constexpr std::size_t kBlockSize = sizeof(V128);
   constexpr std::size_t kCopyRoundSize = kRegions * kBlockSize;
 
   // Number of blocks per cacheline.
@@ -235,7 +238,7 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
   const std::size_t tail_size = length - (kRegions * region_size);
 
   // Holding registers for data in each region.
-  std::array<__m128i, vec_regions> vec_data;
+  std::array<V128, vec_regions> vec_data;
   std::array<uint64_t, int_regions * kIntLoadsPerVec> int_data;
 
   // Main loop.
@@ -243,7 +246,10 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
     // Prefetch kPrefetchAhead bytes ahead of each pointer.
     for (size_t i = 0; i < kRegions; i++) {
       absl::PrefetchToLocalCache(src_bytes + kPrefetchAhead + region_size * i);
+#ifdef ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE
+      // TODO(b/297082454): investigate dropping prefetch on x86.
       absl::PrefetchToLocalCache(dst_bytes + kPrefetchAhead + region_size * i);
+#endif
     }
 
     // Load and store data, computing CRC on the way.
@@ -256,21 +262,20 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
         size_t region = (j + i) % kRegions;
 
         auto* vsrc =
-            reinterpret_cast<const __m128i*>(src_bytes + region_size * region);
-        auto* vdst =
-            reinterpret_cast<__m128i*>(dst_bytes + region_size * region);
+            reinterpret_cast<const V128*>(src_bytes + region_size * region);
+        auto* vdst = reinterpret_cast<V128*>(dst_bytes + region_size * region);
 
         // Load and CRC data.
-        vec_data[j] = _mm_loadu_si128(vsrc + i);
-        crcs[region] = crc32c_t{static_cast<uint32_t>(_mm_crc32_u64(
-            static_cast<uint32_t>(crcs[region]),
-            static_cast<uint64_t>(_mm_extract_epi64(vec_data[j], 0))))};
-        crcs[region] = crc32c_t{static_cast<uint32_t>(_mm_crc32_u64(
-            static_cast<uint32_t>(crcs[region]),
-            static_cast<uint64_t>(_mm_extract_epi64(vec_data[j], 1))))};
+        vec_data[j] = V128_LoadU(vsrc + i);
+        crcs[region] = crc32c_t{static_cast<uint32_t>(
+            CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                      static_cast<uint64_t>(V128_Extract64<0>(vec_data[j]))))};
+        crcs[region] = crc32c_t{static_cast<uint32_t>(
+            CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                      static_cast<uint64_t>(V128_Extract64<1>(vec_data[j]))))};
 
         // Store the data.
-        _mm_store_si128(vdst + i, vec_data[j]);
+        V128_Store(vdst + i, vec_data[j]);
       }
 
       // Preload the partial CRCs for the CLMUL subregions.
@@ -290,7 +295,7 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
 
           // Load and CRC the data.
           int_data[data_index] = *(usrc + i * kIntLoadsPerVec + k);
-          crcs[region] = crc32c_t{static_cast<uint32_t>(_mm_crc32_u64(
+          crcs[region] = crc32c_t{static_cast<uint32_t>(CRC32_u64(
               static_cast<uint32_t>(crcs[region]), int_data[data_index]))};
 
           // Store the data.
@@ -313,6 +318,21 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
   src_bytes += region_size * (kRegions - 1);
   dst_bytes += region_size * (kRegions - 1);
 
+  // Copy and CRC the tail through the XMM registers.
+  std::size_t tail_blocks = tail_size / kBlockSize;
+  LargeTailCopy<0, 1>(&crcs[kRegions - 1], &dst_bytes, &src_bytes, 0,
+                      tail_blocks);
+
+  // Final tail copy for under 16 bytes.
+  crcs[kRegions - 1] =
+      ShortCrcCopy(dst_bytes, src_bytes, tail_size - tail_blocks * kBlockSize,
+                   crcs[kRegions - 1]);
+
+  if (kRegions == 1) {
+    // If there is only one region, finalize and return its CRC.
+    return crc32c_t{static_cast<uint32_t>(crcs[0]) ^ kCrcDataXor};
+  }
+
   // Finalize the first CRCs: XOR the internal CRCs by the XOR mask to undo the
   // XOR done before doing block copy + CRCs.
   for (size_t i = 0; i + 1 < kRegions; i++) {
@@ -325,16 +345,6 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
     full_crc = ConcatCrc32c(full_crc, crcs[i], region_size);
   }
 
-  // Copy and CRC the tail through the XMM registers.
-  std::size_t tail_blocks = tail_size / kBlockSize;
-  LargeTailCopy<0, 1>(&crcs[kRegions - 1], &dst_bytes, &src_bytes, 0,
-                      tail_blocks);
-
-  // Final tail copy for under 16 bytes.
-  crcs[kRegions - 1] =
-      ShortCrcCopy(dst_bytes, src_bytes, tail_size - tail_blocks * kBlockSize,
-                   crcs[kRegions - 1]);
-
   // Finalize and concatenate the final CRC, then return.
   crcs[kRegions - 1] =
       crc32c_t{static_cast<uint32_t>(crcs[kRegions - 1]) ^ kCrcDataXor};
@@ -347,9 +357,11 @@ CrcMemcpy::ArchSpecificEngines CrcMemcpy::GetArchSpecificEngines() {
   // Get the underlying architecture.
   CpuType cpu_type = GetCpuType();
   switch (cpu_type) {
-    case CpuType::kUnknown:
     case CpuType::kAmdRome:
     case CpuType::kAmdNaples:
+    case CpuType::kAmdMilan:
+    case CpuType::kAmdGenoa:
+    case CpuType::kAmdRyzenV3000:
     case CpuType::kIntelCascadelakeXeon:
     case CpuType::kIntelSkylakeXeon:
     case CpuType::kIntelSkylake:
@@ -385,6 +397,9 @@ CrcMemcpy::ArchSpecificEngines CrcMemcpy::GetArchSpecificEngines() {
     // strided access to each region, and do the right thing.
     case CpuType::kAmdRome:
     case CpuType::kAmdNaples:
+    case CpuType::kAmdMilan:
+    case CpuType::kAmdGenoa:
+    case CpuType::kAmdRyzenV3000:
       return {
           /*.temporal=*/new AcceleratedCrcMemcpyEngine<1, 2>(),
           /*.non_temporal=*/new CrcNonTemporalMemcpyAVXEngine(),
@@ -421,6 +436,8 @@ std::unique_ptr<CrcMemcpyEngine> CrcMemcpy::GetTestEngine(int vector,
     return std::make_unique<AcceleratedCrcMemcpyEngine<3, 0>>();
   } else if (vector == 1 && integer == 2) {
     return std::make_unique<AcceleratedCrcMemcpyEngine<1, 2>>();
+  } else if (vector == 1 && integer == 0) {
+    return std::make_unique<AcceleratedCrcMemcpyEngine<1, 0>>();
   }
   return nullptr;
 }
@@ -429,4 +446,5 @@ std::unique_ptr<CrcMemcpyEngine> CrcMemcpy::GetTestEngine(int vector,
 ABSL_NAMESPACE_END
 }  // namespace absl
 
-#endif  // ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE
+#endif  // ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE ||
+        // ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE

data/third_party/abseil-cpp/absl/crc/internal/crc_x86_arm_combined.cc

@@ -16,14 +16,14 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <memory>
+#include <vector>
 
 #include "absl/base/attributes.h"
 #include "absl/base/config.h"
-#include "absl/base/dynamic_annotations.h"
 #include "absl/base/internal/endian.h"
 #include "absl/base/prefetch.h"
 #include "absl/crc/internal/cpu_detect.h"
-#include "absl/crc/internal/crc.h"
 #include "absl/crc/internal/crc32_x86_arm_combined_simd.h"
 #include "absl/crc/internal/crc_internal.h"
 #include "absl/memory/memory.h"
@@ -634,8 +634,16 @@ CRCImpl* TryNewCRC32AcceleratedX86ARMCombined() {
       return new CRC32AcceleratedX86ARMCombinedMultipleStreams<
           3, 0, CutoffStrategy::Fold3>();
     case CpuType::kArmNeoverseN1:
+    case CpuType::kArmNeoverseN2:
+    case CpuType::kArmNeoverseV1:
       return new CRC32AcceleratedX86ARMCombinedMultipleStreams<
           1, 1, CutoffStrategy::Unroll64CRC>();
+    case CpuType::kAmpereSiryn:
+      return new CRC32AcceleratedX86ARMCombinedMultipleStreams<
+          3, 2, CutoffStrategy::Fold3>();
+    case CpuType::kArmNeoverseV2:
+      return new CRC32AcceleratedX86ARMCombinedMultipleStreams<
+          1, 2, CutoffStrategy::Unroll64CRC>();
 #if defined(__aarch64__)
     default:
       // Not all ARM processors support the needed instructions, so check here

data/third_party/abseil-cpp/absl/debugging/internal/address_is_readable.cc

@@ -50,8 +50,10 @@ namespace debugging_internal {
 // NOTE: any new system calls here may also require sandbox reconfiguration.
 //
 bool AddressIsReadable(const void *addr) {
-  // Align address on 8-byte boundary. On aarch64, checking last
-  // byte before inaccessible page returned unexpected EFAULT.
+  // rt_sigprocmask below checks 8 contiguous bytes. If addr resides in the
+  // last 7 bytes of a page (unaligned), rt_sigprocmask would additionally
+  // check the readability of the next page, which is not desired. Align
+  // address on 8-byte boundary to check only the current page.
   const uintptr_t u_addr = reinterpret_cast<uintptr_t>(addr) & ~uintptr_t{7};
   addr = reinterpret_cast<const void *>(u_addr);
 

data/third_party/abseil-cpp/absl/debugging/internal/demangle.cc

@@ -21,7 +21,15 @@
 
 #include <cstdint>
 #include <cstdio>
+#include <cstdlib>
 #include <limits>
+#include <string>
+
+#include "absl/base/config.h"
+
+#if ABSL_INTERNAL_HAS_CXA_DEMANGLE
+#include <cxxabi.h>
+#endif
 
 namespace absl {
 ABSL_NAMESPACE_BEGIN
@@ -1983,6 +1991,22 @@ bool Demangle(const char* mangled, char* out, size_t out_size) {
          state.parse_state.out_cur_idx > 0;
 }
 
+std::string DemangleString(const char* mangled) {
+  std::string out;
+  int status = 0;
+  char* demangled = nullptr;
+#if ABSL_INTERNAL_HAS_CXA_DEMANGLE
+  demangled = abi::__cxa_demangle(mangled, nullptr, nullptr, &status);
+#endif
+  if (status == 0 && demangled != nullptr) {
+    out.append(demangled);
+    free(demangled);
+  } else {
+    out.append(mangled);
+  }
+  return out;
+}
+
 }  // namespace debugging_internal
 ABSL_NAMESPACE_END
 }  // namespace absl

data/third_party/abseil-cpp/absl/debugging/internal/demangle.h

@@ -12,13 +12,23 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// An async-signal-safe and thread-safe demangler for Itanium C++ ABI
-// (aka G++ V3 ABI).
+#ifndef ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
+#define ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
+
+#include <string>
+#include "absl/base/config.h"
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+namespace debugging_internal {
+
+// Demangle `mangled`.  On success, return true and write the
+// demangled symbol name to `out`.  Otherwise, return false.
+// `out` is modified even if demangling is unsuccessful.
 //
-// The demangler is implemented to be used in async signal handlers to
-// symbolize stack traces.  We cannot use libstdc++'s
-// abi::__cxa_demangle() in such signal handlers since it's not async
-// signal safe (it uses malloc() internally).
+// This function provides an alternative to libstdc++'s abi::__cxa_demangle,
+// which is not async signal safe (it uses malloc internally).  It's intended to
+// be used in async signal handlers to symbolize stack traces.
 //
 // Note that this demangler doesn't support full demangling.  More
 // specifically, it doesn't print types of function parameters and
@@ -30,40 +40,32 @@
 //
 // Example:
 //
-// | Mangled Name  | The Demangler | abi::__cxa_demangle()
-// |---------------|---------------|-----------------------
-// | _Z1fv         | f()           | f()
-// | _Z1fi         | f()           | f(int)
-// | _Z3foo3bar    | foo()         | foo(bar)
-// | _Z1fIiEvi     | f<>()         | void f<int>(int)
-// | _ZN1N1fE      | N::f          | N::f
-// | _ZN3Foo3BarEv | Foo::Bar()    | Foo::Bar()
-// | _Zrm1XS_"     | operator%()   | operator%(X, X)
-// | _ZN3FooC1Ev   | Foo::Foo()    | Foo::Foo()
-// | _Z1fSs        | f()           | f(std::basic_string<char,
-// |               |               |   std::char_traits<char>,
-// |               |               |   std::allocator<char> >)
+// | Mangled Name  | Demangle    | DemangleString
+// |---------------|-------------|-----------------------
+// | _Z1fv         | f()         | f()
+// | _Z1fi         | f()         | f(int)
+// | _Z3foo3bar    | foo()       | foo(bar)
+// | _Z1fIiEvi     | f<>()       | void f<int>(int)
+// | _ZN1N1fE      | N::f        | N::f
+// | _ZN3Foo3BarEv | Foo::Bar()  | Foo::Bar()
+// | _Zrm1XS_"     | operator%() | operator%(X, X)
+// | _ZN3FooC1Ev   | Foo::Foo()  | Foo::Foo()
+// | _Z1fSs        | f()         | f(std::basic_string<char,
+// |               |             |   std::char_traits<char>,
+// |               |             |   std::allocator<char> >)
 //
 // See the unit test for more examples.
 //
 // Note: we might want to write demanglers for ABIs other than Itanium
 // C++ ABI in the future.
-//
-
-#ifndef ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
-#define ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Demangle `mangled`.  On success, return true and write the
-// demangled symbol name to `out`.  Otherwise, return false.
-// `out` is modified even if demangling is unsuccessful.
 bool Demangle(const char* mangled, char* out, size_t out_size);
 
+// A wrapper around `abi::__cxa_demangle()`.  On success, returns the demangled
+// name.  On failure, returns the input mangled name.
+//
+// This function is not async-signal-safe.
+std::string DemangleString(const char* mangled);
+
 }  // namespace debugging_internal
 ABSL_NAMESPACE_END
 }  // namespace absl