SingleWriterFixedHashMap: Optimize copying

Summary: Use memcpy for copying maps with the same capacity instead of rehashing.

Reviewed By: yfeldblum

Differential Revision: D20863257

fbshipit-source-id: 11e52f9f2246440fb724a94c2611e44e253e462d

folly/experimental/SingleWriterFixedHashMap.h

@@ -86,6 +86,12 @@ class SingleWriterFixedHashMap {
       return;
     }
     elem_ = std::make_unique<Elem[]>(capacity_);
+    if (capacity_ == o.capacity_) {
+      memcpy(elem_.get(), o.elem_.get(), capacity_ * sizeof(Elem));
+      used_ = o.used_;
+      setSize(o.size());
+      return;
+    }
     for (size_t i = 0; i < o.capacity_; ++i) {
       Elem& e = o.elem_[i];
       if (e.valid()) {

folly/experimental/test/SingleWriterHashMapTest.cpp

@@ -296,6 +296,20 @@ uint64_t bench_copy_nonempty_dtor(const int nthr, const uint64_t ops) {
   return runBench(ops, repFn);
 }
 
+uint64_t bench_copy_resize_dtor(const int nthr, const uint64_t ops) {
+  auto repFn = [&] {
+    SWFHM m0(16);
+    m0.insert(10, 10);
+    auto fn = [&](int tid) {
+      for (uint64_t i = tid; i < ops; i += nthr) {
+        SWFHM m(2 * m0.capacity(), m0);
+      }
+    };
+    return run_once(nthr, fn);
+  };
+  return runBench(ops, repFn);
+}
+
 void dottedLine() {
   std::cout
       << "........................................................................"
@@ -336,6 +350,8 @@ TEST(SingleWriterFixedHashMapBench, Bench) {
     bench_copy_empty_dtor(i, ops);
     std::cout << "copy nonempty / destruct       ";
     bench_copy_nonempty_dtor(i, ops);
+    std::cout << "copy resize / destruct         ";
+    bench_copy_resize_dtor(i, ops);
   }
   std::cout
       << "========================================================================"
@@ -354,12 +370,14 @@ Test name                         Max time  Avg time  Dev time  Min time
 iterate 10-element 32-slot map      20 ns     19 ns      0 ns     19 ns
 construct / destruct                 1 ns      1 ns      0 ns      1 ns
 copy empty / destruct                5 ns      4 ns      0 ns      4 ns
-copy nonempty / destruct            39 ns     38 ns      0 ns     37 ns
+copy nonempty / destruct            25 ns     24 ns      0 ns     23 ns
+copy resize / destruct              40 ns     38 ns      0 ns     37 ns
 ============================== 10 threads ==============================
 10x find                             6 ns      4 ns      1 ns      3 ns
 iterate 10-element 32-slot map       3 ns      3 ns      0 ns      1 ns
 construct / destruct                 0 ns      0 ns      0 ns      0 ns
 copy empty / destruct                0 ns      0 ns      0 ns      0 ns
-copy nonempty / destruct             7 ns      5 ns      1 ns      3 ns
+copy nonempty / destruct             4 ns      3 ns      0 ns      2 ns
+copy resize / destruct               8 ns      6 ns      1 ns      4 ns
 ========================================================================
  */

folly/io/async/test/RequestContextBenchmark.cpp

@@ -300,8 +300,8 @@ onSet                                12 ns     12 ns      0 ns     12 ns
 onUnset                              12 ns     12 ns      0 ns     12 ns
 setContext                           46 ns     44 ns      1 ns     42 ns
 RequestContextScopeGuard            113 ns    103 ns      3 ns    101 ns
-ShallowCopyRequestC...-replace      310 ns    291 ns      8 ns    283 ns
-ShallowCopyReq...-keep&replace     1562 ns   1444 ns     45 ns   1405 ns
+ShallowCopyRequestC...-replace      287 ns    279 ns      5 ns    270 ns
+ShallowCopyReq...-keep&replace     1191 ns   1149 ns     21 ns   1121 ns
 ============================== 10 threads ==============================
 hasContextData                        1 ns      1 ns      0 ns      1 ns
 getContextData                        2 ns      1 ns      0 ns      1 ns
@@ -309,7 +309,7 @@ onSet                                 2 ns      2 ns      0 ns      1 ns
 onUnset                               2 ns      2 ns      0 ns      1 ns
 setContext                           11 ns      7 ns      2 ns      5 ns
 RequestContextScopeGuard             22 ns     15 ns      5 ns     11 ns
-ShallowCopyRequestC...-replace       64 ns     36 ns     10 ns     31 ns
-ShallowCopyReq...-keep&replace      161 ns    156 ns      3 ns    152 ns
+ShallowCopyRequestC...-replace       61 ns     43 ns     14 ns     30 ns
+ShallowCopyReq...-keep&replace      123 ns    121 ns      0 ns    120 ns
 ========================================================================
  */