grpc 1.67.0.pre1 → 1.68.1

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

@@ -33,14 +33,15 @@
 #include <x86intrin.h>
 #define ABSL_CRC_INTERNAL_HAVE_X86_SIMD
 
-#elif defined(_MSC_VER) && !defined(__clang__) && defined(__AVX__)
+#elif defined(_MSC_VER) && !defined(__clang__) && defined(__AVX__) && \
+    defined(_M_AMD64)
 
 // MSVC AVX (/arch:AVX) implies SSE 4.2 and PCLMULQDQ.
 #include <intrin.h>
 #define ABSL_CRC_INTERNAL_HAVE_X86_SIMD
 
-#elif defined(__aarch64__) && defined(__LITTLE_ENDIAN__) && \
-    defined(__ARM_FEATURE_CRC32) && defined(ABSL_INTERNAL_HAVE_ARM_NEON) &&  \
+#elif defined(__aarch64__) && defined(__LITTLE_ENDIAN__) &&                 \
+    defined(__ARM_FEATURE_CRC32) && defined(ABSL_INTERNAL_HAVE_ARM_NEON) && \
     defined(__ARM_FEATURE_CRYPTO)
 
 #include <arm_acle.h>
@@ -101,10 +102,11 @@ V128 V128_Xor(const V128 l, const V128 r);
 // Produces an AND operation of |l| and |r|.
 V128 V128_And(const V128 l, const V128 r);
 
-// Sets two 64 bit integers to one 128 bit vector. The order is reverse.
+// Sets the lower half of a 128 bit register to the given 64-bit value and
+// zeroes the upper half.
 // dst[63:0] := |r|
-// dst[127:64] := |l|
-V128 V128_From2x64(const uint64_t l, const uint64_t r);
+// dst[127:64] := |0|
+V128 V128_From64WithZeroFill(const uint64_t r);
 
 // Shift |l| right by |imm| bytes while shifting in zeros.
 template <int imm>
@@ -121,8 +123,8 @@ uint64_t V128_Extract64(const V128 l);
 // Extracts the low 64 bits from V128.
 int64_t V128_Low64(const V128 l);
 
-// Left-shifts packed 64-bit integers in l by r.
-V128 V128_ShiftLeft64(const V128 l, const V128 r);
+// Add packed 64-bit integers in |l| and |r|.
+V128 V128_Add64(const V128 l, const V128 r);
 
 #endif
 
@@ -170,8 +172,8 @@ inline V128 V128_Xor(const V128 l, const V128 r) { return _mm_xor_si128(l, r); }
 
 inline V128 V128_And(const V128 l, const V128 r) { return _mm_and_si128(l, r); }
 
-inline V128 V128_From2x64(const uint64_t l, const uint64_t r) {
-  return _mm_set_epi64x(static_cast<int64_t>(l), static_cast<int64_t>(r));
+inline V128 V128_From64WithZeroFill(const uint64_t r) {
+  return _mm_set_epi64x(static_cast<int64_t>(0), static_cast<int64_t>(r));
 }
 
 template <int imm>
@@ -191,8 +193,8 @@ inline uint64_t V128_Extract64(const V128 l) {
 
 inline int64_t V128_Low64(const V128 l) { return _mm_cvtsi128_si64(l); }
 
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return _mm_sll_epi64(l, r);
+inline V128 V128_Add64(const V128 l, const V128 r) {
+  return _mm_add_epi64(l, r);
 }
 
 #elif defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
@@ -261,10 +263,12 @@ inline V128 V128_Xor(const V128 l, const V128 r) { return veorq_u64(l, r); }
 
 inline V128 V128_And(const V128 l, const V128 r) { return vandq_u64(l, r); }
 
-inline V128 V128_From2x64(const uint64_t l, const uint64_t r) {
-  return vcombine_u64(vcreate_u64(r), vcreate_u64(l));
+inline V128 V128_From64WithZeroFill(const uint64_t r){
+  constexpr uint64x2_t kZero = {0, 0};
+  return vsetq_lane_u64(r, kZero, 0);
 }
 
+
 template <int imm>
 inline V128 V128_ShiftRight(const V128 l) {
   return vreinterpretq_u64_s8(
@@ -285,9 +289,7 @@ inline int64_t V128_Low64(const V128 l) {
   return vgetq_lane_s64(vreinterpretq_s64_u64(l), 0);
 }
 
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return vshlq_u64(l, vreinterpretq_s64_u64(r));
-}
+inline V128 V128_Add64(const V128 l, const V128 r) { return vaddq_u64(l, r); }
 
 #endif
 

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

@@ -17,6 +17,7 @@
 #include <cassert>
 
 #include "absl/base/config.h"
+#include "absl/base/no_destructor.h"
 #include "absl/numeric/bits.h"
 
 namespace absl {
@@ -24,14 +25,14 @@ ABSL_NAMESPACE_BEGIN
 namespace crc_internal {
 
 CrcCordState::RefcountedRep* CrcCordState::RefSharedEmptyRep() {
-  static CrcCordState::RefcountedRep* empty = new CrcCordState::RefcountedRep;
+  static absl::NoDestructor<CrcCordState::RefcountedRep> empty;
 
   assert(empty->count.load(std::memory_order_relaxed) >= 1);
   assert(empty->rep.removed_prefix.length == 0);
   assert(empty->rep.prefix_crc.empty());
 
-  Ref(empty);
-  return empty;
+  Ref(empty.get());
+  return empty.get();
 }
 
 CrcCordState::CrcCordState() : refcounted_rep_(new RefcountedRep) {}

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

@@ -12,12 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <cstdint>
+#include <cstring>
 #include <memory>
 
 #include "absl/base/config.h"
 #include "absl/crc/crc32c.h"
 #include "absl/crc/internal/crc_memcpy.h"
+#include "absl/strings/string_view.h"
 
 namespace absl {
 ABSL_NAMESPACE_BEGIN

data/third_party/abseil-cpp/absl/crc/internal/crc_memcpy_x86_arm_combined.cc

@@ -52,6 +52,7 @@
 #include <cstring>
 #include <memory>
 
+#include "absl/base/attributes.h"
 #include "absl/base/config.h"
 #include "absl/base/optimization.h"
 #include "absl/base/prefetch.h"
@@ -88,9 +89,11 @@ inline crc32c_t ShortCrcCopy(char* dst, const char* src, std::size_t length,
 constexpr size_t kIntLoadsPerVec = sizeof(V128) / sizeof(uint64_t);
 
 // Common function for copying the tails of multiple large regions.
+// Disable ubsan for benign unaligned access. See b/254108538.
 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) {
+ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED inline void LargeTailCopy(
+    crc32c_t* crcs, char** dst, const char** src, size_t region_size,
+    size_t copy_rounds) {
   std::array<V128, vec_regions> data;
   std::array<uint64_t, kIntLoadsPerVec * int_regions> int_data;
 
@@ -127,8 +130,8 @@ 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>(CRC32_u64(
-            static_cast<uint32_t>(crcs[region]), int_data[data_index]))};
+        crcs[region] = crc32c_t{CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                                          int_data[data_index])};
 
         *(udst + j) = int_data[data_index];
       }
@@ -155,8 +158,10 @@ class AcceleratedCrcMemcpyEngine : public CrcMemcpyEngine {
                    std::size_t length, crc32c_t initial_crc) const override;
 };
 
+// Disable ubsan for benign unaligned access. See b/254108538.
 template <size_t vec_regions, size_t int_regions>
-crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
+ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED 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;
@@ -196,7 +201,6 @@ crc32c_t AcceleratedCrcMemcpyEngine<vec_regions, int_regions>::Compute(
 
   // Start work on the CRC: undo the XOR from the previous calculation or set up
   // the initial value of the CRC.
-  // initial_crc ^= kCrcDataXor;
   initial_crc = crc32c_t{static_cast<uint32_t>(initial_crc) ^ kCrcDataXor};
 
   // Do an initial alignment copy, so we can use aligned store instructions to
@@ -295,8 +299,8 @@ 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>(CRC32_u64(
-              static_cast<uint32_t>(crcs[region]), int_data[data_index]))};
+          crcs[region] = crc32c_t{CRC32_u64(static_cast<uint32_t>(crcs[region]),
+                                            int_data[data_index])};
 
           // Store the data.
           *(udst + i * kIntLoadsPerVec + k) = int_data[data_index];

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

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <cstdint>
+#include <cstddef>
 
 #include "absl/base/config.h"
 #include "absl/crc/crc32c.h"

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

@@ -101,13 +101,17 @@ constexpr size_t kMediumCutoff = 2048;
 namespace {
 
 uint32_t multiply(uint32_t a, uint32_t b) {
-  V128 shifts = V128_From2x64(0, 1);
-  V128 power = V128_From2x64(0, a);
-  V128 crc = V128_From2x64(0, b);
+  V128 power = V128_From64WithZeroFill(a);
+  V128 crc = V128_From64WithZeroFill(b);
   V128 res = V128_PMulLow(power, crc);
 
-  // Combine crc values
-  res = V128_ShiftLeft64(res, shifts);
+  // Combine crc values.
+  //
+  // Adding res to itself is equivalent to multiplying by 2,
+  // or shifting left by 1. Addition is used as not all compilers
+  // are able to generate optimal code without this hint.
+  // https://godbolt.org/z/rr3fMnf39
+  res = V128_Add64(res, res);
   return static_cast<uint32_t>(V128_Extract32<1>(res)) ^
          CRC32_u32(0, static_cast<uint32_t>(V128_Low64(res)));
 }
@@ -444,11 +448,11 @@ class CRC32AcceleratedX86ARMCombinedMultipleStreams
 
         V128 magic = *(reinterpret_cast<const V128*>(kClmulConstants) + bs - 1);
 
-        V128 tmp = V128_From2x64(0, l64);
+        V128 tmp = V128_From64WithZeroFill(l64);
 
         V128 res1 = V128_PMulLow(tmp, magic);
 
-        tmp = V128_From2x64(0, l641);
+        tmp = V128_From64WithZeroFill(l641);
 
         V128 res2 = V128_PMul10(tmp, magic);
         V128 x = V128_Xor(res1, res2);

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

@@ -19,19 +19,8 @@
 #include <intrin.h>
 #endif
 
-#ifdef __SSE__
-#include <xmmintrin.h>
-#endif
-
-#ifdef __SSE2__
-#include <emmintrin.h>
-#endif
-
-#ifdef __SSE3__
-#include <pmmintrin.h>
-#endif
-
-#ifdef __AVX__
+#if defined(__SSE__) || defined(__AVX__)
+// Pulls in both SSE and AVX intrinsics.
 #include <immintrin.h>
 #endif
 
@@ -44,6 +33,7 @@
 #include <cstdint>
 #include <cstring>
 
+#include "absl/base/attributes.h"
 #include "absl/base/config.h"
 #include "absl/base/optimization.h"
 
@@ -57,7 +47,9 @@ namespace crc_internal {
 // memcpy can save 1 DRAM load of the destination cacheline.
 constexpr size_t kCacheLineSize = ABSL_CACHELINE_SIZE;
 
-// If the objects overlap, the behavior is undefined.
+// If the objects overlap, the behavior is undefined. Uses regular memcpy
+// instead of non-temporal memcpy if the required CPU intrinsics are unavailable
+// at compile time.
 inline void *non_temporal_store_memcpy(void *__restrict dst,
                                        const void *__restrict src, size_t len) {
 #if defined(__SSE3__) || defined(__aarch64__) || \
@@ -119,10 +111,20 @@ inline void *non_temporal_store_memcpy(void *__restrict dst,
 #endif  // __SSE3__ || __aarch64__ || (_MSC_VER && __AVX__)
 }
 
+// If the objects overlap, the behavior is undefined. Uses regular memcpy
+// instead of non-temporal memcpy if the required CPU intrinsics are unavailable
+// at compile time.
+#if ABSL_HAVE_CPP_ATTRIBUTE(gnu::target) && \
+    (defined(__x86_64__) || defined(__i386__))
+[[gnu::target("avx")]]
+#endif
 inline void *non_temporal_store_memcpy_avx(void *__restrict dst,
                                            const void *__restrict src,
                                            size_t len) {
-#ifdef __AVX__
+  // This function requires AVX. For clang and gcc we compile it with AVX even
+  // if the translation unit isn't built with AVX support. This works because we
+  // only select this implementation at runtime if the CPU supports AVX.
+#if defined(__SSE3__) || (defined(_MSC_VER) && defined(__AVX__))
   uint8_t *d = reinterpret_cast<uint8_t *>(dst);
   const uint8_t *s = reinterpret_cast<const uint8_t *>(src);
 
@@ -168,9 +170,8 @@ inline void *non_temporal_store_memcpy_avx(void *__restrict dst,
   }
   return dst;
 #else
-  // Fallback to regular memcpy when AVX is not available.
   return memcpy(dst, src, len);
-#endif  // __AVX__
+#endif  // __SSE3__ || (_MSC_VER && __AVX__)
 }
 
 }  // namespace crc_internal

data/third_party/abseil-cpp/absl/debugging/internal/bounded_utf8_length_sequence.h

@@ -0,0 +1,126 @@
+// Copyright 2024 The Abseil Authors
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ABSL_DEBUGGING_INTERNAL_BOUNDED_UTF8_LENGTH_SEQUENCE_H_
+#define ABSL_DEBUGGING_INTERNAL_BOUNDED_UTF8_LENGTH_SEQUENCE_H_
+
+#include <cstdint>
+
+#include "absl/base/config.h"
+#include "absl/numeric/bits.h"
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+namespace debugging_internal {
+
+// A sequence of up to max_elements integers between 1 and 4 inclusive, whose
+// insertion operation computes the sum of all the elements before the insertion
+// point.  This is useful in decoding Punycode, where one needs to know where in
+// a UTF-8 byte stream the n-th code point begins.
+//
+// BoundedUtf8LengthSequence is async-signal-safe and suitable for use in
+// symbolizing stack traces in a signal handler, provided max_elements is not
+// improvidently large.  For inputs of lengths accepted by the Rust demangler,
+// up to a couple hundred code points, InsertAndReturnSumOfPredecessors should
+// run in a few dozen clock cycles, on par with the other arithmetic required
+// for Punycode decoding.
+template <uint32_t max_elements>
+class BoundedUtf8LengthSequence {
+ public:
+  // Constructs an empty sequence.
+  BoundedUtf8LengthSequence() = default;
+
+  // Inserts `utf_length` at position `index`, shifting any existing elements at
+  // or beyond `index` one position to the right.  If the sequence is already
+  // full, the rightmost element is discarded.
+  //
+  // Returns the sum of the elements at positions 0 to `index - 1` inclusive.
+  // If `index` is greater than the number of elements already inserted, the
+  // excess positions in the range count 1 apiece.
+  //
+  // REQUIRES: index < max_elements and 1 <= utf8_length <= 4.
+  uint32_t InsertAndReturnSumOfPredecessors(
+      uint32_t index, uint32_t utf8_length) {
+    // The caller shouldn't pass out-of-bounds inputs, but if it does happen,
+    // clamp the values and try to continue.  If we're being called from a
+    // signal handler, the last thing we want to do is crash.  Emitting
+    // malformed UTF-8 is a lesser evil.
+    if (index >= max_elements) index = max_elements - 1;
+    if (utf8_length == 0 || utf8_length > 4) utf8_length = 1;
+
+    const uint32_t word_index = index/32;
+    const uint32_t bit_index = 2 * (index % 32);
+    const uint64_t ones_bit = uint64_t{1} << bit_index;
+
+    // Compute the sum of predecessors.
+    //   - Each value from 1 to 4 is represented by a bit field with value from
+    //     0 to 3, so the desired sum is index plus the sum of the
+    //     representations actually stored.
+    //   - For each bit field, a set low bit should contribute 1 to the sum, and
+    //     a set high bit should contribute 2.
+    //   - Another way to say the same thing is that each set bit contributes 1,
+    //     and each set high bit contributes an additional 1.
+    //   - So the sum we want is index + popcount(everything) + popcount(bits in
+    //     odd positions).
+    const uint64_t odd_bits_mask = 0xaaaaaaaaaaaaaaaa;
+    const uint64_t lower_seminibbles_mask = ones_bit - 1;
+    const uint64_t higher_seminibbles_mask = ~lower_seminibbles_mask;
+    const uint64_t same_word_bits_below_insertion =
+        rep_[word_index] & lower_seminibbles_mask;
+    int full_popcount = absl::popcount(same_word_bits_below_insertion);
+    int odd_popcount =
+        absl::popcount(same_word_bits_below_insertion & odd_bits_mask);
+    for (uint32_t j = word_index; j > 0; --j) {
+      const uint64_t word_below_insertion = rep_[j - 1];
+      full_popcount += absl::popcount(word_below_insertion);
+      odd_popcount += absl::popcount(word_below_insertion & odd_bits_mask);
+    }
+    const uint32_t sum_of_predecessors =
+        index + static_cast<uint32_t>(full_popcount + odd_popcount);
+
+    // Now insert utf8_length's representation, shifting successors up one
+    // place.
+    for (uint32_t j = max_elements/32 - 1; j > word_index; --j) {
+      rep_[j] = (rep_[j] << 2) | (rep_[j - 1] >> 62);
+    }
+    rep_[word_index] =
+        (rep_[word_index] & lower_seminibbles_mask) |
+        (uint64_t{utf8_length - 1} << bit_index) |
+        ((rep_[word_index] & higher_seminibbles_mask) << 2);
+
+    return sum_of_predecessors;
+  }
+
+ private:
+  // If the (32 * i + j)-th element of the represented sequence has the value k
+  // (0 <= j < 32, 1 <= k <= 4), then bits 2 * j and 2 * j + 1 of rep_[i]
+  // contain the seminibble (k - 1).
+  //
+  // In particular, the zero-initialization of rep_ makes positions not holding
+  // any inserted element count as 1 in InsertAndReturnSumOfPredecessors.
+  //
+  // Example: rep_ = {0xb1, ... the rest zeroes ...} represents the sequence
+  // (2, 1, 4, 3, ... the rest 1's ...).  Constructing the sequence of Unicode
+  // code points "Àa🂻中" = {U+00C0, U+0061, U+1F0BB, U+4E2D} (among many
+  // other examples) would yield this value of rep_.
+  static_assert(max_elements > 0 && max_elements % 32 == 0,
+                "max_elements must be a positive multiple of 32");
+  uint64_t rep_[max_elements/32] = {};
+};
+
+}  // namespace debugging_internal
+ABSL_NAMESPACE_END
+}  // namespace absl
+
+#endif  // ABSL_DEBUGGING_INTERNAL_BOUNDED_UTF8_LENGTH_SEQUENCE_H_