ARM: branchless binary search for smallPrefixes_

Summary:
Extract a branchless upper_bound into a function and use it on ARM for the
smallPrefixes_ lookup. Conditional index updates compile to CSEL instructions,
avoiding branch mispredictions that hurt sorted_vector_map::upper_bound()
which goes through iterator/comparator wrappers.

On x86, the hardware branch predictor handles upper_bound's branches well,
so this optimization is gated with `#ifdef __aarch64__`.

Common code (fullPrefixes_ upper_bound and previousPrefix_ walk) is shared
between both paths.

Benchmark command:
```
buck run fbcode/mode/opt fbcode//mcrouter/lib/fbi/cpp/test/facebook:string_prefix_map_benchmark -- --bm_regex _Lb --bm_mode=adaptive --bm_min_secs=5 --bm_max_secs=30
```

# ARM

Before:
```
============================================================================
[...]facebook/StringPrefixMapBenchmark.cpp     relative  time/iter   iters/s
============================================================================
PrefixMapInCache_Lb                                       197.37ns     5.07M
PrefixMapNotInCache_Lb                                      1.83us   547.10K
```

After:
```
============================================================================
[...]facebook/StringPrefixMapBenchmark.cpp     relative  time/iter   iters/s
============================================================================
PrefixMapInCache_Lb                                        92.27ns    10.84M
PrefixMapNotInCache_Lb                                    449.13ns     2.23M
```

- PrefixMapInCache_Lb:    197.37ns → 92.27ns  (53% less time/iter)
- PrefixMapNotInCache_Lb:   1.83us → 449.13ns (75% less time/iter)

Reviewed By: DenisYaroshevskiy

Differential Revision: D97136185

fbshipit-source-id: 2c06d420378264b595a03efab56358857bc30fcc

Author: Alfredo Altamirano

mcrouter/lib/fbi/cpp/LowerBoundPrefixMap.cpp

@@ -13,12 +13,37 @@
 namespace facebook::memcache::detail {
 namespace {
 
-// .starts_with is not avaliable before C++20 and this needs to compile with an
+// .starts_with is not available before C++20 and this needs to compile with an
 // older standard
 bool std_string_view_starts_with(std::string_view what, std::string_view with) {
   return what.substr(0, with.size()) == with;
 }
 
+#ifdef __aarch64__
+// Branchless upper_bound on ARM. Conditional index updates compile to CSEL
+// instructions, avoiding branch mispredictions that hurt
+// sorted_vector_map::upper_bound() through iterator/comparator wrappers.
+// On x86, the hardware branch predictor handles upper_bound well so this
+// is not beneficial there.
+template <typename I, typename T, typename Compare>
+I branchlessUpperBound(I f, I l, const T& x, Compare comp) {
+  auto len = static_cast<std::size_t>(l - f);
+  if (len == 0) {
+    return f;
+  }
+
+  while (len > 1) {
+    auto half = len / 2;
+    f += comp(x, f[half]) ? 0 : half;
+    len -= half;
+  }
+  f += comp(x, *f) ? 0 : 1;
+
+  return f;
+}
+
+#endif
+
 } // namespace
 
 std::ostream& operator<<(std::ostream& os, const SmallPrefix& self) {
@@ -65,8 +90,16 @@ LowerBoundPrefixMapCommon::LowerBoundPrefixMapCommon(
 
 std::uint32_t LowerBoundPrefixMapCommon::findPrefix(
     std::string_view query) const noexcept {
+#ifdef __aarch64__
+  auto afterPrefix = branchlessUpperBound(
+      smallPrefixes_.begin(),
+      smallPrefixes_.end(),
+      SmallPrefix{query},
+      [](const SmallPrefix& v, const auto& elem) { return v < elem.first; });
+#else
   // Due to a sentinel - guaranteed to not be .begin()
   auto afterPrefix = smallPrefixes_.upper_bound(SmallPrefix{query});
+#endif
   auto [roughFrom, roughTo] = std::prev(afterPrefix)->second;
 
   // Binary search complete strings between rough boundaries.