Basic AVX-512 implementation (icelake or better) (#1813)

* Add cascadelake implementation, which use AVX512 Intrinsics to optimize performance(#1811)

Co-authored-by: mellonyou <fangzheng.zhang@intel.com>
Co-authored-by: wanweiqiangintel <weiqiang.wan@intel.com>

.drone.yml

@@ -8,7 +8,7 @@ steps:
     CC: gcc
     CXX: g++
     BUILD_FLAGS: -- -j
-    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=haswell;westmere;fallback
+    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=icelake;haswell;westmere;fallback
     CTEST_FLAGS: -j4 --output-on-failure -LE explicitonly
   commands:
     - echo "deb http://deb.debian.org/debian buster-backports main" >> /etc/apt/sources.list
@@ -19,6 +19,7 @@ steps:
     - cmake $CMAKE_FLAGS ..
     - cmake --build . $BUILD_FLAGS
     - ctest $CTEST_FLAGS -L acceptance -LE per_implementation
+    - SIMDJSON_FORCE_IMPLEMENTATION=icelake ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=haswell ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=westmere ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=fallback ctest $CTEST_FLAGS -L per_implementation
@@ -35,7 +36,7 @@ steps:
     CC: clang-6.0
     CXX: clang++-6.0
     BUILD_FLAGS: -- -j
-    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=haswell;westmere;fallback
+    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=icelake;haswell;westmere;fallback
     CTEST_FLAGS: -j4 --output-on-failure -LE explicitonly
   commands:
     - mkdir build
@@ -43,6 +44,7 @@ steps:
     - cmake $CMAKE_FLAGS ..
     - cmake --build . $BUILD_FLAGS
     - ctest $CTEST_FLAGS -L acceptance -LE per_implementation
+    - SIMDJSON_FORCE_IMPLEMENTATION=icelake ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=haswell ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=westmere ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=fallback ctest $CTEST_FLAGS -L per_implementation
@@ -100,7 +102,7 @@ steps:
     CC: gcc
     CXX: g++
     BUILD_FLAGS: -- -j
-    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=haswell;westmere;fallback
+    CMAKE_FLAGS: -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_IMPLEMENTATION=icelake;haswell;westmere;fallback
     CTEST_FLAGS: -j4 --output-on-failure -LE explicitonly
   commands:
     - echo "deb http://deb.debian.org/debian buster-backports main" >> /etc/apt/sources.list
@@ -111,6 +113,7 @@ steps:
     - cmake $CMAKE_FLAGS ..
     - cmake --build . $BUILD_FLAGS
     - ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L acceptance -LE per_implementation
+    - SIMDJSON_FORCE_IMPLEMENTATION=icelake ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=haswell ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=westmere ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=fallback ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
@@ -126,7 +129,7 @@ steps:
   environment:
     CC: clang-9
     CXX: clang++-9
-    CMAKE_FLAGS: -DSIMDJSON_SANITIZE=ON -DSIMDJSON_IMPLEMENTATION=haswell;westmere;fallback
+    CMAKE_FLAGS: -DSIMDJSON_SANITIZE=ON -DSIMDJSON_IMPLEMENTATION=icelake;haswell;westmere;fallback
     BUILD_FLAGS: -- -j
     CTEST_FLAGS: -j4 --output-on-failure -LE explicitonly
   commands:
@@ -135,6 +138,7 @@ steps:
     - cmake $CMAKE_FLAGS ..
     - cmake --build . $BUILD_FLAGS
     - ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L acceptance -LE per_implementation
+    - SIMDJSON_FORCE_IMPLEMENTATION=icelake ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=haswell ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=westmere ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation
     - SIMDJSON_FORCE_IMPLEMENTATION=fallback ASAN_OPTIONS="detect_leaks=0" ctest $CTEST_FLAGS -L per_implementation

.github/workflows/msys2-clang.yml

@@ -38,6 +38,6 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake -DSIMDJSON_DEVELOPER_MODE=ON -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=${{ matrix.type }} -DSIMDJSON_BUILD_STATIC=ON -DSIMDJSON_DO_NOT_USE_THREADS_NO_MATTER_WHAT=ON ..
+          cmake -DSIMDJSON_DEVELOPER_MODE=ON -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=${{ matrix.type }} -DBUILD_SHARED_LIBS=OFF -DSIMDJSON_DO_NOT_USE_THREADS_NO_MATTER_WHAT=ON ..
           cmake --build . --verbose
           ctest -j4 --output-on-failure -LE explicitonly

.github/workflows/ubuntu22.yml

@@ -0,0 +1,33 @@
+name: Ubuntu 22.04 CI (GCC 11)
+
+on: [push, pull_request]
+
+jobs:
+  ubuntu-build:
+    if: >-
+      ! contains(toJSON(github.event.commits.*.message), '[skip ci]') &&
+      ! contains(toJSON(github.event.commits.*.message), '[skip github]')
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v2
+      - uses: actions/cache@v2
+        with:
+          path: dependencies/.cache
+          key: ${{ hashFiles('dependencies/CMakeLists.txt') }}
+      - name: Use cmake
+        run: |
+          mkdir builddebug &&
+          cd builddebug &&
+          cmake -DCMAKE_BUILD_TYPE=Debug -DSIMDJSON_GOOGLE_BENCHMARKS=OFF -DSIMDJSON_DEVELOPER_MODE=ON -DBUILD_SHARED_LIBS=OFF ..  &&
+          cmake --build .   &&
+          ctest -j --output-on-failure -LE explicitonly   &&
+          cd .. &&
+          mkdir build &&
+          cd build &&
+          cmake  -DSIMDJSON_GOOGLE_BENCHMARKS=ON -DSIMDJSON_DEVELOPER_MODE=ON -DBUILD_SHARED_LIBS=OFF -DCMAKE_INSTALL_PREFIX:PATH=destination ..  &&
+          cmake --build .   &&
+          ctest -j --output-on-failure -LE explicitonly   &&
+          cmake --install . &&
+          echo -e '#include <simdjson.h>\nint main(int argc,char**argv) {simdjson::dom::parser parser;simdjson::dom::element tweets = parser.load(argv[1]); }' > tmp.cpp && c++ -Idestination/include -Ldestination/lib -std=c++17 -Wl,-rpath,destination/lib -o linkandrun tmp.cpp -lsimdjson && ./linkandrun jsonexamples/twitter.json &&
+          cd ../tests/installation_tests/find &&
+          mkdir build && cd build && cmake -DCMAKE_INSTALL_PREFIX:PATH=../../../build/destination .. &&  cmake --build .

.github/workflows/vs17-ci.yml

@@ -13,16 +13,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - {gen: Visual Studio 17 2022, arch: Win32, static: ON}
-          - {gen: Visual Studio 17 2022, arch: Win32, static: OFF}
-          - {gen: Visual Studio 17 2022, arch: x64, static: ON}
-          - {gen: Visual Studio 17 2022, arch: x64, static: OFF}
+          - {gen: Visual Studio 17 2022, arch: Win32, shared: ON}
+          - {gen: Visual Studio 17 2022, arch: Win32, shared: OFF}
+          - {gen: Visual Studio 17 2022, arch: x64, shared: ON}
+          - {gen: Visual Studio 17 2022, arch: x64, shared: OFF}
     steps:
     - name: checkout
       uses: actions/checkout@v2
     - name: Configure
       run: |
-        cmake -G "${{matrix.gen}}" -A ${{matrix.arch}} -DSIMDJSON_DEVELOPER_MODE=ON -DSIMDJSON_COMPETITION=OFF -DSIMDJSON_BUILD_STATIC=${{matrix.static}} -B build
+        cmake -G "${{matrix.gen}}" -A ${{matrix.arch}} -DSIMDJSON_DEVELOPER_MODE=ON -DSIMDJSON_COMPETITION=OFF -DBUILD_SHARED_LIBS=${{matrix.shared}} -B build
     - name: Build Debug
       run: cmake --build build --config Debug --verbose
     - name: Build Release

.github/workflows/vs17-clang-ci.yml

@@ -9,23 +9,33 @@ jobs:
       ! contains(toJSON(github.event.commits.*.message), '[skip github]')
     name: windows-vs17
     runs-on: windows-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - {gen: Visual Studio 17 2022, arch: x64}
     steps:
-    - uses: actions/checkout@v2
-    - uses: actions/cache@v2
-      with:
-        path: dependencies/.cache
-        key: ${{ hashFiles('dependencies/CMakeLists.txt') }}
-    - name: 'Run CMake with VS17 Clang'
-      uses: lukka/run-cmake@v3
-      with:
-        cmakeListsOrSettingsJson: CMakeListsTxtAdvanced
-        cmakeListsTxtPath: '${{ github.workspace }}/CMakeLists.txt'
-        buildDirectory: "${{ github.workspace }}/../../_temp/windows"
-        cmakeBuildType: Release
-        buildWithCMake: true
-        cmakeAppendedArgs:  -T ClangCL -DSIMDJSON_COMPETITION=OFF -DSIMDJSON_DEVELOPER_MODE=ON -DBUILD_SHARED_LIBS=OFF
-        buildWithCMakeArgs: --config Release
-
-    - name: 'Run CTest'
-      run: ctest -C Release -LE explicitonly  --output-on-failure
-      working-directory: "${{ github.workspace }}/../../_temp/windows"
+    - name: checkout
+      uses: actions/checkout@v2
+    - name: Configure
+      run: |
+        cmake -G "${{matrix.gen}}" -A ${{matrix.arch}}  -T ClangCL -DSIMDJSON_DEVELOPER_MODE=ON -DSIMDJSON_COMPETITION=OFF -B build
+    - name: Build Debug
+      run: cmake --build build --config Debug --verbose
+    - name: Build Release
+      run: cmake --build build --config Release --verbose
+    - name: Run Release tests
+      run: |
+        cd build
+        ctest -C Release  -LE explicitonly  --output-on-failure
+    - name: Run Debug tests
+      run: |
+        cd build
+        ctest -C Debug  -LE explicitonly  --output-on-failure
+    - name: Install
+      run: |
+        cmake --install build --config Release
+    - name: Test Installation
+      run: |
+        cmake -G "${{matrix.gen}}" -A ${{matrix.arch}}  -B build_install_test tests/installation_tests/find
+        cmake --build build_install_test --config Release

.github/workflows/vs17-ninja-ci.yml

@@ -1,45 +0,0 @@
-name: VS17-Ninja-CI
-
-on: [push, pull_request]
-
-jobs:
-  ci:
-    if: >-
-      ! contains(toJSON(github.event.commits.*.message), '[skip ci]') &&
-      ! contains(toJSON(github.event.commits.*.message), '[skip github]')
-    name: windows-vs17
-    runs-on: windows-latest
-    steps:
-    - uses: actions/checkout@v2
-    - uses: actions/cache@v2
-      with:
-        path: dependencies/.cache
-        key: ${{ hashFiles('dependencies/CMakeLists.txt') }}
-    - name: 'Run CMake with VS17'
-      uses: lukka/run-cmake@v2
-      with:
-        cmakeListsOrSettingsJson: CMakeListsTxtAdvanced
-        cmakeListsTxtPath: '${{ github.workspace }}/CMakeLists.txt'
-        buildDirectory: "${{ github.workspace }}/../../_temp/windows"
-        cmakeBuildType: Release
-        buildWithCMake: true
-        cmakeAppendedArgs:  -G Ninja -DSIMDJSON_COMPETITION=OFF -DSIMDJSON_DEVELOPER_MODE=ON -DBUILD_SHARED_LIBS=OFF
-        buildWithCMakeArgs: --config Release
-    - name: 'Run CTest'
-      run: ctest -C Release  -LE explicitonly  --output-on-failure
-      working-directory: "${{ github.workspace }}/../../_temp/windows"
-    - name: 'Install with CMake'
-      uses: lukka/run-cmake@v3
-      with:
-        cmakeListsTxtPath: '${{ github.workspace }}/CMakeLists.txt'
-        buildWithCMakeArgs: '--target install'
-    - name: 'Test Installation with CMake'
-      uses: lukka/run-cmake@v3
-      with:
-        cmakeListsOrSettingsJson: CMakeListsTxtAdvanced
-        cmakeListsTxtPath: '${{ github.workspace }}/tests/installation_tests/find/CMakeLists.txt'
-        cmakeBuildType: Release
-        buildWithCMake: true
-        buildDirectory: '${{ github.workspace }}/tests/installation_tests/find/buildDirectory'
-        cmakeAppendedArgs:  -G Ninja
-        buildWithCMakeArgs: '--config Release --verbose'

CONTRIBUTORS

@@ -38,5 +38,8 @@ Brendan Knapp
 Danila Kutenin
 Pavel Pavlov
 Hao Chen
-# if you have contributed to the project and your name does not 
+Nicolas Boyer
+Kim Walisch and Jatin Bhateja (AVX-512 bitset decoder)
+Eric Zhang (AVX-512 kernel)
+# if you have contributed to the project and your name does not
 # appear in this list, please let us know!

benchmark/dom/statisticalmodel.cpp

@@ -50,9 +50,9 @@ simdjson_really_inline void simdjson_process_atom(stat_t &s,
   } else if(element.is<double>()) {
     s.float_count++;
   } else if (element.is<bool>()) {
-    simdjson::error_code err;
     bool v;
-    err = element.get(v);
+    simdjson::error_code error;
+    if ((error = element.get(v))) { std::cerr << error << std::endl; abort(); }
     if (v) {
       s.true_count++;
     } else {

cmake/developer-options.cmake

@@ -175,6 +175,18 @@ if(CMAKE_C_COMPILER_ID MATCHES "Intel")
   set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -static-intel")
 endif()
 
+option(
+    SIMDJSON_AVX512_ALLOWED
+    "Enable AVX-512 instructions (only affects processors and compilers with AVX-512 support)."
+    ON
+)
+if(SIMDJSON_AVX512_ALLOWED)
+  add_compile_definitions(SIMDJSON_AVX512_ALLOWED=1)
+else()
+  add_compile_definitions(SIMDJSON_AVX512_ALLOWED=0)
+  message(STATUS "AVX-512 instructions are not allowed.")
+endif()
+
 include(CheckSymbolExists)
 check_symbol_exists(fork unistd.h HAVE_POSIX_FORK)
 check_symbol_exists(wait sys/wait.h HAVE_POSIX_WAIT)

cmake/implementation-flags.cmake

@@ -1,7 +1,7 @@
 #
 # Implementation selection
 #
-set(SIMDJSON_ALL_IMPLEMENTATIONS fallback westmere haswell arm64 ppc64)
+set(SIMDJSON_ALL_IMPLEMENTATIONS fallback westmere haswell icelake arm64 ppc64)
 
 set(
     SIMDJSON_IMPLEMENTATION ""
@@ -15,7 +15,7 @@ set(
     SIMDJSON_EXCLUDE_IMPLEMENTATION ""
     CACHE STRING "\
 Semicolon-separated list of implementations to exclude \
-(haswell/westmere/arm64/ppc64/fallback). By default, excludes any \
+(icelake/haswell/westmere/arm64/ppc64/fallback). By default, excludes any \
 implementations that are unsupported at compile time or cannot be selected at \
 runtime."
 )

doc/performance.md

@@ -160,6 +160,8 @@ You should not expect the simdjson library to cause *downclocking* of your recen
 - [Whenever 512-bit AVX-512 instructions are used](https://lemire.me/blog/2018/09/07/avx-512-when-and-how-to-use-these-new-instructions/).
 - Whenever heavy 256-bit or wider instructions are used. Heavy instructions are those involving floating point operations or integer multiplications (since these execute on the floating point unit).
 
-The simdjson library does not currently support AVX-512 instructions and it does not make use of heavy 256-bit instructions. We do use vectorized multiplications, but only using 128-bit registers. Thus there should be no downclocking due to simdjson on recent processors.
+The simdjson library does not make use of heavy 256-bit instructions. We do use vectorized multiplications, but only using 128-bit registers. Thus there should be no downclocking due to simdjson on recent processors, except when AVX-512 is allowed and
+detected. However, we only allow AVX-512 on recent processors (Ice Lake/Tiger Lake or better) where [little to no frequency throttling is expected](https://travisdowns.github.io/blog/2020/08/19/icl-avx512-freq.html). If you can still concerned, you can easily disable AVX-512 with the CMake option `SIMDJSON_AVX512_ALLOWED` set to `OFF` (e.g., `cmake -D SIMDJSON_AVX512_ALLOWED=OFF -B build && cmake --build build`) or by setting
+the macro `SIMDJSON_AVX512_ALLOWED` to `0` in C++ prior to importing the headers.
 
 You may still be worried about which SIMD instruction set is used by simdjson.  Thankfully,  [you can always determine and change which architecture-specific implementation is used](implementation-selection.md) by simdjson. Thus even if your CPU supports AVX2, you do not need to use AVX2. You are in control.

include/simdjson/builtin.h

@@ -5,7 +5,9 @@
 
 // Determine the best builtin implementation
 #ifndef SIMDJSON_BUILTIN_IMPLEMENTATION
-#if SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+#if SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+#define SIMDJSON_BUILTIN_IMPLEMENTATION icelake
+#elif SIMDJSON_CAN_ALWAYS_RUN_HASWELL
 #define SIMDJSON_BUILTIN_IMPLEMENTATION haswell
 #elif SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
 #define SIMDJSON_BUILTIN_IMPLEMENTATION westmere

include/simdjson/haswell/intrinsics.h

@@ -41,11 +41,7 @@
 // has it as a macro.
 #ifndef _blsr_u64
 // we roll our own
-SIMDJSON_TARGET_HASWELL
-static simdjson_really_inline uint64_t _blsr_u64(uint64_t n) {
-  return (n - 1) & n;
-}
-SIMDJSON_UNTARGET_HASWELL
+#define _blsr_u64(n) ((n - 1) & n)
 #endif //  _blsr_u64
 #endif // SIMDJSON_CLANG_VISUAL_STUDIO
 

include/simdjson/icelake.h

@@ -0,0 +1,47 @@
+#ifndef SIMDJSON_ICELAKE_H
+#define SIMDJSON_ICELAKE_H
+
+#include "simdjson/implementation-base.h"
+
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+
+#if SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+#define SIMDJSON_TARGET_ICELAKE
+#define SIMDJSON_UNTARGET_ICELAKE
+#else
+#define SIMDJSON_TARGET_ICELAKE SIMDJSON_TARGET_REGION("avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,avx512vl,avx2,bmi,pclmul,lzcnt")
+#define SIMDJSON_UNTARGET_ICELAKE SIMDJSON_UNTARGET_REGION
+#endif
+
+namespace simdjson {
+/**
+ * Implementation for Icelake (Intel AVX512).
+ */
+namespace icelake {
+} // namespace icelake
+} // namespace simdjson
+
+//
+// These two need to be included outside SIMDJSON_TARGET_ICELAKE
+//
+#include "simdjson/icelake/implementation.h"
+#include "simdjson/icelake/intrinsics.h"
+
+//
+// The rest need to be inside the region
+//
+#include "simdjson/icelake/begin.h"
+
+// Declarations
+#include "simdjson/generic/dom_parser_implementation.h"
+#include "simdjson/icelake/bitmanipulation.h"
+#include "simdjson/icelake/bitmask.h"
+#include "simdjson/icelake/simd.h"
+#include "simdjson/generic/jsoncharutils.h"
+#include "simdjson/generic/atomparsing.h"
+#include "simdjson/icelake/stringparsing.h"
+#include "simdjson/icelake/numberparsing.h"
+#include "simdjson/icelake/end.h"
+
+#endif // SIMDJSON_IMPLEMENTATION_ICELAKE
+#endif // SIMDJSON_ICELAKE_H

include/simdjson/icelake/begin.h

@@ -0,0 +1,2 @@
+#define SIMDJSON_IMPLEMENTATION icelake
+SIMDJSON_TARGET_ICELAKE

include/simdjson/icelake/bitmanipulation.h

@@ -0,0 +1,61 @@
+#ifndef SIMDJSON_ICELAKE_BITMANIPULATION_H
+#define SIMDJSON_ICELAKE_BITMANIPULATION_H
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+#ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  ////////
+  // You might expect the next line to be equivalent to
+  // return (int)_tzcnt_u64(input_num);
+  // but the generated code differs and might be less efficient?
+  ////////
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return _blsr_u64(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+  return int(_lzcnt_u64(input_num));
+}
+
+#ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_really_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMANIPULATION_H

include/simdjson/icelake/bitmask.h

@@ -0,0 +1,25 @@
+#ifndef SIMDJSON_ICELAKE_BITMASK_H
+#define SIMDJSON_ICELAKE_BITMASK_H
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_really_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processor supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMASK_H

include/simdjson/icelake/end.h

@@ -0,0 +1,2 @@
+SIMDJSON_UNTARGET_ICELAKE
+#undef SIMDJSON_IMPLEMENTATION

include/simdjson/icelake/implementation.h

@@ -0,0 +1,31 @@
+#ifndef SIMDJSON_ICELAKE_IMPLEMENTATION_H
+#define SIMDJSON_ICELAKE_IMPLEMENTATION_H
+
+#include "simdjson/implementation.h"
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_ICELAKE
+namespace simdjson {
+namespace icelake {
+
+using namespace simdjson;
+
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_really_inline implementation() : simdjson::implementation(
+      "icelake",
+      "Intel/AMD AVX512",
+      internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2 | internal::instruction_set::AVX512F | internal::instruction_set::AVX512DQ | internal::instruction_set::AVX512CD | internal::instruction_set::AVX512BW | internal::instruction_set::AVX512VL | internal::instruction_set::AVX512VBMI2
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_IMPLEMENTATION_H

include/simdjson/icelake/intrinsics.h

@@ -0,0 +1,56 @@
+#ifndef SIMDJSON_ICELAKE_INTRINSICS_H
+#define SIMDJSON_ICELAKE_INTRINSICS_H
+
+#include "simdjson/base.h"
+
+#ifdef SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h>  // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+#ifdef SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ * e.g., if __AVX2__ is set... in turn,  we normally set these
+ * macros by compiling against the corresponding architecture
+ * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+ * software with these advanced instructions. In simdjson, we
+ * want to compile the whole program for a generic target,
+ * and only target our specific kernels. As a workaround,
+ * we directly include the needed headers. These headers would
+ * normally guard against such usage, but we carefully included
+ * <x86intrin.h>  (or <intrin.h>) before, so the headers
+ * are fooled.
+ */
+#include <bmiintrin.h>   // for _blsr_u64
+#include <lzcntintrin.h> // for  __lzcnt64
+#include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+#include <smmintrin.h>
+#include <tmmintrin.h>
+#include <avxintrin.h>
+#include <avx2intrin.h>
+#include <wmmintrin.h>   // for  _mm_clmulepi64_si128
+// Important: we need the AVX-512 headers:
+#include <avx512fintrin.h>
+#include <avx512dqintrin.h>
+#include <avx512cdintrin.h>
+#include <avx512bwintrin.h>
+#include <avx512vlintrin.h>
+#include <avx512vbmiintrin.h>
+#include <avx512vbmi2intrin.h>
+// unfortunately, we may not get _blsr_u64, but, thankfully, clang
+// has it as a macro.
+#ifndef _blsr_u64
+// we roll our own
+#define _blsr_u64(n) ((n - 1) & n)
+#endif //  _blsr_u64
+#endif // SIMDJSON_CLANG_VISUAL_STUDIO
+
+#endif // SIMDJSON_ICELAKE_INTRINSICS_H

include/simdjson/icelake/numberparsing.h

@@ -0,0 +1,34 @@
+#ifndef SIMDJSON_ICELAKE_NUMBERPARSING_H
+#define SIMDJSON_ICELAKE_NUMBERPARSING_H
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+
+static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#include "simdjson/generic/numberparsing.h"
+
+#endif // SIMDJSON_ICELAKE_NUMBERPARSING_H

include/simdjson/icelake/simd.h

@@ -0,0 +1,339 @@
+#ifndef SIMDJSON_ICELAKE_SIMD_H
+#define SIMDJSON_ICELAKE_SIMD_H
+
+#include "simdjson/internal/simdprune_tables.h"
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m512i value;
+
+    // Zero constructor
+    simdjson_really_inline base() : value{__m512i()} {}
+
+    // Conversion from SIMD register
+    simdjson_really_inline base(const __m512i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_really_inline operator const __m512i&() const { return this->value; }
+    simdjson_really_inline operator __m512i&() { return this->value; }
+
+    // Bit operations
+    simdjson_really_inline Child operator|(const Child other) const { return _mm512_or_si512(*this, other); }
+    simdjson_really_inline Child operator&(const Child other) const { return _mm512_and_si512(*this, other); }
+    simdjson_really_inline Child operator^(const Child other) const { return _mm512_xor_si512(*this, other); }
+    simdjson_really_inline Child bit_andnot(const Child other) const { return _mm512_andnot_si512(other, *this); }
+    simdjson_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_really_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint32_t bitmask_t;
+    typedef uint64_t bitmask2_t;
+
+    simdjson_really_inline base8() : base<simd8<T>>() {}
+    simdjson_really_inline base8(const __m512i _value) : base<simd8<T>>(_value) {}
+
+    simdjson_really_inline uint64_t operator==(const simd8<T> other) const { return _mm512_cmpeq_epi8_mask(*this, other); }
+
+    static const int SIZE = sizeof(base<T>::value);
+
+    template<int N=1>
+    simdjson_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm512_alignr_epi8(*this, _mm512_permutex2var_epi64(prev_chunk, _mm512_set_epi64(13, 12, 11, 10, 9, 8, 7, 6), *this), 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_really_inline simd8<bool> splat(bool _value) { return _mm512_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_really_inline simd8<bool>() : base8() {}
+    simdjson_really_inline simd8<bool>(const __m512i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_really_inline bool any() const { return !!_mm512_test_epi8_mask (*this, *this); }
+    simdjson_really_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_really_inline simd8<T> splat(T _value) { return _mm512_set1_epi8(_value); }
+    static simdjson_really_inline simd8<T> zero() { return _mm512_setzero_si512(); }
+    static simdjson_really_inline simd8<T> load(const T values[64]) {
+      return _mm512_loadu_si512(reinterpret_cast<const __m512i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_really_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_really_inline base8_numeric() : base8<T>() {}
+    simdjson_really_inline base8_numeric(const __m512i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_really_inline void store(T dst[64]) const { return _mm512_storeu_si512(reinterpret_cast<__m512i *>(dst), *this); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm512_add_epi8(*this, other); }
+    simdjson_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm512_sub_epi8(*this, other); }
+    simdjson_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_really_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm512_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 32 - count_ones(mask) bytes of the result are significant but 32 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_really_inline void compress(uint64_t mask, L * output) const {
+      _mm512_mask_compressstoreu_epi8 (output,~mask,*this);
+    }
+
+    template<typename L>
+    simdjson_really_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_really_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_really_inline simd8(const __m512i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_really_inline simd8(const int8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_really_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31,
+      int8_t v32, int8_t v33, int8_t v34, int8_t v35, int8_t v36, int8_t v37, int8_t v38, int8_t v39,
+      int8_t v40, int8_t v41, int8_t v42, int8_t v43, int8_t v44, int8_t v45, int8_t v46, int8_t v47,
+      int8_t v48, int8_t v49, int8_t v50, int8_t v51, int8_t v52, int8_t v53, int8_t v54, int8_t v55,
+      int8_t v56, int8_t v57, int8_t v58, int8_t v59, int8_t v60, int8_t v61, int8_t v62, int8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_really_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm512_max_epi8(*this, other); }
+    simdjson_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm512_min_epi8(*this, other); }
+
+    simdjson_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(*this, other),_mm512_set1_epi8(uint8_t(0x80))); }
+    simdjson_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(other, *this),_mm512_set1_epi8(uint8_t(0x80))); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_really_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_really_inline simd8(const __m512i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_really_inline simd8(const uint8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_really_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31,
+      uint8_t v32, uint8_t v33, uint8_t v34, uint8_t v35, uint8_t v36, uint8_t v37, uint8_t v38, uint8_t v39,
+      uint8_t v40, uint8_t v41, uint8_t v42, uint8_t v43, uint8_t v44, uint8_t v45, uint8_t v46, uint8_t v47,
+      uint8_t v48, uint8_t v49, uint8_t v50, uint8_t v51, uint8_t v52, uint8_t v53, uint8_t v54, uint8_t v55,
+      uint8_t v56, uint8_t v57, uint8_t v58, uint8_t v59, uint8_t v60, uint8_t v61, uint8_t v62, uint8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_really_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm512_adds_epu8(*this, other); }
+    simdjson_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm512_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm512_max_epu8(*this, other); }
+    simdjson_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm512_min_epu8(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_really_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_really_inline uint64_t operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_really_inline uint64_t operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_really_inline simd8<bool> bits_not_set() const { return _mm512_mask_blend_epi8(*this == uint8_t(0), _mm512_set1_epi8(0), _mm512_set1_epi8(-1)); }
+    simdjson_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+
+    simdjson_really_inline bool is_ascii() const { return _mm512_movepi8_mask(*this) == 0; }
+    simdjson_really_inline bool bits_not_set_anywhere() const {
+      return !_mm512_test_epi8_mask(*this, *this);
+    }
+    simdjson_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return !_mm512_test_epi8_mask(*this, bits); }
+    simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm512_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_really_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm512_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_really_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 1, "Icelake kernel should use one register per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_really_inline simd8x64(const simd8<T> chunk0) : chunks{chunk0} {}
+    simdjson_really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr)} {}
+
+    simdjson_really_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(mask, output);
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_really_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+    }
+
+    simdjson_really_inline simd8<T> reduce_or() const {
+      return this->chunks[0];
+    }
+
+    simdjson_really_inline simd8x64<T> bit_or(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return simd8x64<T>(
+        this->chunks[0] | mask
+      );
+    }
+
+    simdjson_really_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] == mask;
+    }
+
+    simdjson_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return this->chunks[0] == other.chunks[0];
+    }
+
+    simdjson_really_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] <= mask;
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_SIMD_H

include/simdjson/icelake/stringparsing.h

@@ -0,0 +1,48 @@
+#ifndef SIMDJSON_ICELAKE_STRINGPARSING_H
+#define SIMDJSON_ICELAKE_STRINGPARSING_H
+
+#include "simdjson/base.h"
+#include "simdjson/icelake/simd.h"
+#include "simdjson/icelake/bitmanipulation.h"
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_really_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_really_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_really_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_really_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint64_t bs_bits;
+  uint64_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 15 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  v.store(dst);
+  return {
+      static_cast<uint64_t>(v == '\\'), // bs_bits
+      static_cast<uint64_t>(v == '"'), // quote_bits
+  };
+}
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#include "simdjson/generic/stringparsing.h"
+
+#endif // SIMDJSON_ICELAKE_STRINGPARSING_H

include/simdjson/implementations.h

@@ -13,10 +13,32 @@
 #endif
 #define SIMDJSON_CAN_ALWAYS_RUN_ARM64 SIMDJSON_IMPLEMENTATION_ARM64 && SIMDJSON_IS_ARM64
 
+#ifdef __has_include
+// How do we detect that a compiler supports vbmi2?
+// For sure if the following header is found, we are ok?
+#if __has_include(<avx512vbmi2intrin.h>)
+#define SIMDJSON_COMPILER_SUPPORTS_VBMI2 1
+#endif
+#endif
+
+// Default Icelake to on if this is x86-64. Even if we're not compiled for it, it could be selected
+// at runtime.
+#ifndef SIMDJSON_IMPLEMENTATION_ICELAKE
+#define SIMDJSON_IMPLEMENTATION_ICELAKE ((SIMDJSON_IS_X86_64) && (SIMDJSON_AVX512_ALLOWED) && (SIMDJSON_COMPILER_SUPPORTS_VBMI2))
+#endif
+
+#ifdef _MSC_VER
+// To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdjson/simdjson/issues/1247
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE ((SIMDJSON_IMPLEMENTATION_ICELAKE) && (__AVX2__) && (__AVX512F__) && (__AVX512DQ__) && (__AVX512CD__) && (__AVX512BW__) && (__AVX512VL__) && (__AVX512VBMI2__))
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE ((SIMDJSON_IMPLEMENTATION_ICELAKE) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__) && (__AVX512F__) && (__AVX512DQ__) && (__AVX512CD__) && (__AVX512BW__) && (__AVX512VL__) && (__AVX512VBMI2__))
+#endif
+
 // Default Haswell to on if this is x86-64. Even if we're not compiled for it, it could be selected
 // at runtime.
 #ifndef SIMDJSON_IMPLEMENTATION_HASWELL
-#define SIMDJSON_IMPLEMENTATION_HASWELL (SIMDJSON_IS_X86_64)
+#define SIMDJSON_IMPLEMENTATION_HASWELL SIMDJSON_IS_X86_64
 #endif
 #ifdef _MSC_VER
 // To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
@@ -26,7 +48,7 @@
 #define SIMDJSON_CAN_ALWAYS_RUN_HASWELL ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))
 #endif
 
-// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
+// Default Westmere to on if this is x86-64. Note that the macro SIMDJSON_REQUIRES_HASWELL appears unused.
 #ifndef SIMDJSON_IMPLEMENTATION_WESTMERE
 #define SIMDJSON_IMPLEMENTATION_WESTMERE (SIMDJSON_IS_X86_64 && !SIMDJSON_REQUIRES_HASWELL)
 #endif
@@ -49,6 +71,7 @@ SIMDJSON_DISABLE_UNDESIRED_WARNINGS
 // Implementations
 #include "simdjson/arm64.h"
 #include "simdjson/fallback.h"
+#include "simdjson/icelake.h"
 #include "simdjson/haswell.h"
 #include "simdjson/ppc64.h"
 #include "simdjson/westmere.h"

include/simdjson/internal/isadetection.h

@@ -66,7 +66,16 @@ enum instruction_set {
   PCLMULQDQ = 0x10,
   BMI1 = 0x20,
   BMI2 = 0x40,
-  ALTIVEC = 0x80
+  ALTIVEC = 0x80,
+  AVX512F = 0x100,
+  AVX512DQ = 0x200,
+  AVX512IFMA = 0x400,
+  AVX512PF = 0x800,
+  AVX512ER = 0x1000,
+  AVX512CD = 0x2000,
+  AVX512BW = 0x4000,
+  AVX512VL = 0x8000,
+  AVX512VBMI2 = 0x10000
 };
 
 #if defined(__PPC64__)
@@ -96,11 +105,20 @@ static inline uint32_t detect_supported_architectures() {
 
 namespace {
 // Can be found on Intel ISA Reference for CPUID
-constexpr uint32_t cpuid_avx2_bit = 1 << 5;      ///< @private Bit 5 of EBX for EAX=0x7
-constexpr uint32_t cpuid_bmi1_bit = 1 << 3;      ///< @private bit 3 of EBX for EAX=0x7
-constexpr uint32_t cpuid_bmi2_bit = 1 << 8;      ///< @private bit 8 of EBX for EAX=0x7
-constexpr uint32_t cpuid_sse42_bit = 1 << 20;    ///< @private bit 20 of ECX for EAX=0x1
-constexpr uint32_t cpuid_pclmulqdq_bit = 1 << 1; ///< @private bit  1 of ECX for EAX=0x1
+constexpr uint32_t cpuid_avx2_bit = 1 << 5;         ///< @private Bit 5 of EBX for EAX=0x7
+constexpr uint32_t cpuid_bmi1_bit = 1 << 3;         ///< @private bit 3 of EBX for EAX=0x7
+constexpr uint32_t cpuid_bmi2_bit = 1 << 8;         ///< @private bit 8 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512f_bit = 1 << 16;     ///< @private bit 16 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512dq_bit = 1 << 17;    ///< @private bit 17 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512ifma_bit = 1 << 21;  ///< @private bit 21 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512pf_bit = 1 << 26;    ///< @private bit 26 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512er_bit = 1 << 27;    ///< @private bit 27 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512cd_bit = 1 << 28;    ///< @private bit 28 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512bw_bit = 1 << 30;    ///< @private bit 30 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512vl_bit = 1 << 31;    ///< @private bit 31 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512vbmi2_bit = 1 << 6;  ///< @private bit 6 of ECX for EAX=0x7
+constexpr uint32_t cpuid_sse42_bit = 1 << 20;       ///< @private bit 20 of ECX for EAX=0x1
+constexpr uint32_t cpuid_pclmulqdq_bit = 1 << 1;    ///< @private bit  1 of ECX for EAX=0x1
 }
 
 
@@ -146,6 +164,42 @@ static inline uint32_t detect_supported_architectures() {
     host_isa |= instruction_set::BMI2;
   }
 
+  if (ebx & cpuid_avx512f_bit) {
+    host_isa |= instruction_set::AVX512F;
+  }
+
+  if (ebx & cpuid_avx512dq_bit) {
+    host_isa |= instruction_set::AVX512DQ;
+  }
+
+  if (ebx & cpuid_avx512ifma_bit) {
+    host_isa |= instruction_set::AVX512IFMA;
+  }
+
+  if (ebx & cpuid_avx512pf_bit) {
+    host_isa |= instruction_set::AVX512PF;
+  }
+
+  if (ebx & cpuid_avx512er_bit) {
+    host_isa |= instruction_set::AVX512ER;
+  }
+
+  if (ebx & cpuid_avx512cd_bit) {
+    host_isa |= instruction_set::AVX512CD;
+  }
+
+  if (ebx & cpuid_avx512bw_bit) {
+    host_isa |= instruction_set::AVX512BW;
+  }
+
+  if (ebx & cpuid_avx512vl_bit) {
+    host_isa |= instruction_set::AVX512VL;
+  }
+
+  if (ecx & cpuid_avx512vbmi2_bit) {
+    host_isa |= instruction_set::AVX512VBMI2;
+  }
+
   // EBX for EAX=0x1
   eax = 0x1;
   cpuid(&eax, &ebx, &ecx, &edx);

src/generic/stage1/json_structural_indexer.h

@@ -25,6 +25,12 @@ public:
   // base_ptr[base] incrementing base as we go
   // will potentially store extra values beyond end of valid bits, so base_ptr
   // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
+  // version of the code.
+#ifdef SIMDJSON_CUSTOM_BIT_INDEXER
+  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+#else
   simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
@@ -117,6 +123,8 @@ public:
     this->tail += cnt;
 #endif
   }
+#endif // SIMDJSON_CUSTOM_BIT_INDEXER
+
 };
 
 class json_structural_indexer {

src/generic/stage1/utf8_lookup4_algorithm.h

@@ -111,12 +111,25 @@ using namespace simd;
   simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+    };
+#else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
     };
+#endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
     return input.gt_bits(max_value);
   }
@@ -154,9 +167,13 @@ using namespace simd;
         this->error |= this->prev_incomplete;
       } else {
         // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
-        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
         } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {

src/icelake/dom_parser_implementation.cpp

@@ -0,0 +1,206 @@
+#include "simdjson/icelake/begin.h"
+
+//
+// Stage 1
+//
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+
+using namespace simd;
+
+struct json_character_block {
+  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  //  ASCII white-space ('\r','\n','\t',' ')
+  simdjson_really_inline uint64_t whitespace() const noexcept;
+  // non-quote structural characters (comma, colon, braces, brackets)
+  simdjson_really_inline uint64_t op() const noexcept;
+  // neither a structural character nor a white-space, so letters, numbers and quotes
+  simdjson_really_inline uint64_t scalar() const noexcept;
+
+  uint64_t _whitespace; // ASCII white-space ('\r','\n','\t',' ')
+  uint64_t _op; // structural characters (comma, colon, braces, brackets but not quotes)
+};
+
+simdjson_really_inline uint64_t json_character_block::whitespace() const noexcept { return _whitespace; }
+simdjson_really_inline uint64_t json_character_block::op() const noexcept { return _op; }
+simdjson_really_inline uint64_t json_character_block::scalar() const noexcept { return ~(op() | whitespace()); }
+
+// This identifies structural characters (comma, colon, braces, brackets),
+// and ASCII white-space ('\r','\n','\t',' ').
+simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
+  // we can't use the generic lookup_16.
+  const auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
+
+  // The 6 operators (:,[]{}) have these values:
+  //
+  // , 2C
+  // : 3A
+  // [ 5B
+  // { 7B
+  // ] 5D
+  // } 7D
+  //
+  // If you use | 0x20 to turn [ and ] into { and }, the lower 4 bits of each character is unique.
+  // We exploit this, using a simd 4-bit lookup to tell us which character match against, and then
+  // match it (against | 0x20).
+  //
+  // To prevent recognizing other characters, everything else gets compared with 0, which cannot
+  // match due to the | 0x20.
+  //
+  // NOTE: Due to the | 0x20, this ALSO treats <FF> and <SUB> (control characters 0C and 1A) like ,
+  // and :. This gets caught in stage 2, which checks the actual character to ensure the right
+  // operators are in the right places.
+  const auto op_table = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0,
+    0, 0, 0, 0,
+    0, 0, ':', '{', // : = 3A, [ = 5B, { = 7B
+    ',', '}', 0, 0  // , = 2C, ] = 5D, } = 7D
+  );
+
+  // We compute whitespace and op separately. If later code only uses one or the
+  // other, given the fact that all functions are aggressively inlined, we can
+  // hope that useless computations will be omitted. This is namely case when
+  // minifying (we only need whitespace).
+
+  const uint64_t whitespace = in.eq({
+    _mm512_shuffle_epi8(whitespace_table, in.chunks[0])
+  });
+  // Turn [ and ] into { and }
+  const simd8x64<uint8_t> curlified{
+    in.chunks[0] | 0x20
+  };
+  const uint64_t op = curlified.eq({
+    _mm512_shuffle_epi8(op_table, in.chunks[0])
+  });
+
+  return { whitespace, op };
+}
+
+simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0b11000000u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+} // unnamed namespace
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#include "generic/stage1/utf8_lookup4_algorithm.h"
+// defining SIMDJSON_CUSTOM_BIT_INDEXER allows us to provide our own bit_indexer::write
+#define SIMDJSON_CUSTOM_BIT_INDEXER
+#include "generic/stage1/json_structural_indexer.h"
+// We must not forget to undefine it now:
+#undef SIMDJSON_CUSTOM_BIT_INDEXER
+
+/**
+ * We provide a custom version of bit_indexer::write using
+ * naked intrinsics.
+ * TODO: make this code more elegant.
+ */
+namespace simdjson { namespace SIMDJSON_IMPLEMENTATION { namespace { namespace stage1 {
+simdjson_really_inline void bit_indexer::write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0) { return; }
+
+    const __m512i indexes = _mm512_maskz_compress_epi8(bits, _mm512_set_epi32(
+      0x3f3e3d3c, 0x3b3a3938, 0x37363534, 0x33323130,
+      0x2f2e2d2c, 0x2b2a2928, 0x27262524, 0x23222120,
+      0x1f1e1d1c, 0x1b1a1918, 0x17161514, 0x13121110,
+      0x0f0e0d0c, 0x0b0a0908, 0x07060504, 0x03020100
+    ));
+    const __m512i start_index = _mm512_set1_epi32(idx);
+
+    const auto count = count_ones(bits);
+    __m512i t0 = _mm512_cvtepu8_epi32(_mm512_castsi512_si128(indexes));
+    _mm512_storeu_si512(this->tail, _mm512_add_epi32(t0, start_index));
+
+    if(count > 16) {
+      const __m512i t1 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 1));
+      _mm512_storeu_si512(this->tail + 16, _mm512_add_epi32(t1, start_index));
+      if(count > 32) {
+        const __m512i t2 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 2));
+        _mm512_storeu_si512(this->tail + 32, _mm512_add_epi32(t2, start_index));
+        if(count > 48) {
+          const __m512i t3 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 3));
+          _mm512_storeu_si512(this->tail + 48, _mm512_add_epi32(t3, start_index));
+        }
+      }
+    }
+    this->tail += count;
+}
+}}}}
+
+#include "generic/stage1/utf8_validator.h"
+
+//
+// Stage 2
+//
+#include "generic/stage2/tape_builder.h"
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+namespace {
+namespace stage1 {
+
+simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+  if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
+  return find_escaped_branchless(backslash);
+}
+
+} // namespace stage1
+} // unnamed namespace
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return icelake::stage1::json_minifier::minify<128>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return icelake::stage1::json_structural_indexer::index<128>(_buf, _len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return icelake::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#include "simdjson/icelake/end.h"

src/icelake/implementation.cpp

@@ -0,0 +1,24 @@
+#include "simdjson/icelake/begin.h"
+
+namespace simdjson {
+namespace SIMDJSON_IMPLEMENTATION {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+} // namespace SIMDJSON_IMPLEMENTATION
+} // namespace simdjson
+
+#include "simdjson/icelake/end.h"
+

src/implementation.cpp

@@ -14,6 +14,12 @@ namespace internal {
 // Static array of known implementations. We're hoping these get baked into the executable
 // without requiring a static initializer.
 
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+static const icelake::implementation* get_icelake_singleton() {
+  static const icelake::implementation icelake_singleton{};
+  return &icelake_singleton;
+}
+#endif
 #if SIMDJSON_IMPLEMENTATION_HASWELL
 static const haswell::implementation* get_haswell_singleton() {
   static const haswell::implementation haswell_singleton{};
@@ -73,6 +79,9 @@ private:
 
 static const std::initializer_list<const implementation *>& get_available_implementation_pointers() {
   static const std::initializer_list<const implementation *> available_implementation_pointers {
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+    get_icelake_singleton(),
+#endif
 #if SIMDJSON_IMPLEMENTATION_HASWELL
     get_haswell_singleton(),
 #endif

src/internal/simdprune_tables.cpp

@@ -1,4 +1,4 @@
-#if SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64
+#if SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_ICELAKE || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64
 
 #include <cstdint>
 
@@ -128,4 +128,4 @@ SIMDJSON_DLLIMPORTEXPORT  const uint64_t thintable_epi8[256] = {
 } // namespace internal
 } // namespace simdjson
 
-#endif //  SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64
+#endif //  SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_ICELAKE || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64

src/simdjson.cpp

@@ -19,6 +19,10 @@ SIMDJSON_DISABLE_UNDESIRED_WARNINGS
 #include "fallback/implementation.cpp"
 #include "fallback/dom_parser_implementation.cpp"
 #endif
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+#include "icelake/implementation.cpp"
+#include "icelake/dom_parser_implementation.cpp"
+#endif
 #if SIMDJSON_IMPLEMENTATION_HASWELL
 #include "haswell/implementation.cpp"
 #include "haswell/dom_parser_implementation.cpp"

tests/ondemand/ondemand_array_error_tests.cpp

@@ -16,7 +16,7 @@ namespace array_error_tests {
     size_t count = 0;
     for (auto elem : std::forward<T>(array)) {
       std::cout << "-"; std::cout.flush();
-      V actual;
+      V actual{};
       auto actual_error = elem.get(actual);
       if (count >= N) {
         if (count >= (N+N2)) {

tests/ondemand/ondemand_readme_examples.cpp

@@ -752,7 +752,7 @@ bool stream_capacity_example() {
   if( error ) { /* handle the error */ }
   for (auto doc: stream) {
     if(counter < 6) {
-      int64_t val;
+      int64_t val{};
       error = doc.at_pointer("/4").get(val);
       if( error ) { /* handle the error */ }
       std::cout << "5 = " << val << std::endl;