Add multi-version rezero16

ConflictSet.cpp

@@ -1332,6 +1332,29 @@ void maybeDecreaseCapacity(Node *&self, WriteContext *tls,
   freeAndMakeCapacityAtLeast(self, maxCapacity, tls, impl, false);
 }
 
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+// This gets covered in local development
+// GCOVR_EXCL_START
+__attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
+                                                 InternalVersionT zero) {
+  uint32_t z;
+  memcpy(&z, &zero, sizeof(z));
+  const auto zvec = _mm512_set1_epi32(z);
+  auto m = _mm512_cmplt_epi32_mask(
+      _mm512_sub_epi32(_mm512_loadu_epi32(vs), zvec), _mm512_setzero_epi32());
+  _mm512_mask_storeu_epi32(vs, m, zvec);
+}
+// GCOVR_EXCL_STOP
+
+__attribute__((target("default")))
+#endif
+
+void rezero16(InternalVersionT *vs, InternalVersionT zero) {
+  for (int i = 0; i < 16; ++i) {
+    vs[i] = std::max(vs[i], zero);
+  }
+}
+
 void rezero(Node *n, InternalVersionT z) {
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "rezero to %" PRId64 ": %s\n", z.toInt64(),
@@ -1352,14 +1375,12 @@ void rezero(Node *n, InternalVersionT z) {
   } break;
   case Type_Node16: {
     auto *self = static_cast<Node16 *>(n);
-    for (int i = 0; i < 16; ++i) {
-      self->childMaxVersion[i] = std::max(self->childMaxVersion[i], z);
-    }
+    rezero16(self->childMaxVersion, z);
   } break;
   case Type_Node48: {
     auto *self = static_cast<Node48 *>(n);
-    for (int i = 0; i < 48; ++i) {
-      self->childMaxVersion[i] = std::max(self->childMaxVersion[i], z);
+    for (int i = 0; i < 48; i += 16) {
+      rezero16(self->childMaxVersion + i, z);
     }
     for (auto &m : self->maxOfMax) {
       m = std::max(m, z);
@@ -1367,12 +1388,10 @@ void rezero(Node *n, InternalVersionT z) {
   } break;
   case Type_Node256: {
     auto *self = static_cast<Node256 *>(n);
-    for (int i = 0; i < 256; ++i) {
-      self->childMaxVersion[i] = std::max(self->childMaxVersion[i], z);
-    }
-    for (auto &m : self->maxOfMax) {
-      m = std::max(m, z);
+    for (int i = 0; i < 256; i += 16) {
+      rezero16(self->childMaxVersion + i, z);
     }
+    rezero16(self->maxOfMax, z);
   } break;
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE