Add USE_SIMD_FALLBACK

Closes #13

CMakeLists.txt

@@ -25,6 +25,9 @@ endif()
 add_compile_options(-fdata-sections -ffunction-sections -Wswitch-enum
                     -Werror=switch-enum)
 
+option(USE_SIMD_FALLBACK
+       "Use fallback implementations of functions that use SIMD" OFF)
+
 # This is encouraged according to
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
 include_directories(SYSTEM ${CMAKE_SOURCE_DIR}/third_party/valgrind)
@@ -43,6 +46,7 @@ endif()
 include(CheckIncludeFileCXX)
 include(CMakePushCheckState)
 
+if(NOT USE_SIMD_FALLBACK)
   cmake_push_check_state()
   list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
   check_include_file_cxx("immintrin.h" HAS_AVX)
@@ -56,6 +60,7 @@ check_include_file_cxx("arm_neon.h" HAS_ARM_NEON)
   if(HAS_ARM_NEON)
     add_compile_definitions(HAS_ARM_NEON)
   endif()
+endif()
 
 set(CMAKE_CXX_IMPLICIT_LINK_LIBRARIES "")
 

ConflictSet.cpp

@@ -29,6 +29,7 @@ limitations under the License.
 #include <span>
 #include <string>
 #include <string_view>
+#include <type_traits>
 #include <utility>
 
 #ifdef HAS_AVX
@@ -954,6 +955,42 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     assert(self->getType() == Type_Node16);
 
     ++self->numChildren;
+#ifdef HAS_AVX
+    __m128i key_vec = _mm_set1_epi8(index);
+    __m128i indices;
+    memcpy(&indices, self16->index, sizeof(self16->index));
+    __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
+    int mask = (1 << (self->numChildren - 1)) - 1;
+    uint32_t bitfield = _mm_movemask_epi8(results) & mask;
+    bitfield |= uint32_t(1) << (self->numChildren - 1);
+    int i = std::countr_zero(bitfield);
+    if (i < self->numChildren - 1) {
+      memmove(self16->index + i + 1, self16->index + i,
+              self->numChildren - (i + 1));
+      memmove(self16->children + i + 1, self16->children + i,
+              (self->numChildren - (i + 1)) * sizeof(Child));
+    }
+#elif defined(HAS_ARM_NEON)
+    uint8x16_t indices;
+    memcpy(&indices, self16->index, sizeof(self16->index));
+    // 0xff for each leq
+    auto results = vcleq_u8(vdupq_n_u8(index), indices);
+    uint64_t mask = (uint64_t(1) << ((self->numChildren - 1) * 4)) - 1;
+    // 0xf for each 0xff (within mask)
+    uint64_t bitfield =
+        vget_lane_u64(
+            vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(results), 4)),
+            0) &
+        mask;
+    bitfield |= uint64_t(0xf) << ((self->numChildren - 1) * 4);
+    int i = std::countr_zero(bitfield) / 4;
+    if (i < self->numChildren - 1) {
+      memmove(self16->index + i + 1, self16->index + i,
+              self->numChildren - (i + 1));
+      memmove(self16->children + i + 1, self16->children + i,
+              (self->numChildren - (i + 1)) * sizeof(Child));
+    }
+#else
     int i = 0;
     for (; i < int(self->numChildren) - 1; ++i) {
       if (int(self16->index[i]) > int(index)) {
@@ -964,6 +1001,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
         break;
       }
     }
+#endif
     self16->index[i] = index;
     auto &result = self16->children[i].child;
     result = nullptr;
@@ -979,8 +1017,8 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       self = newSelf;
       goto insert256;
     }
-  insert48:
 
+  insert48:
     auto *self48 = static_cast<Node48 *>(self);
     self48->bitSet.set(index);
     ++self->numChildren;
@@ -992,6 +1030,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     return result;
   }
   case Type_Node256: {
+
   insert256:
     auto *self256 = static_cast<Node256 *>(self);
     ++self->numChildren;