crlib: replace random number generator with wyhash.

Benchmarks shows about 4x increased PRNG perfomance increase comparing to old one. Should slightly reduce per-bot CPU usage.

ext/crlib/cr-array.h

@@ -272,8 +272,10 @@ public:
    }
 
    void shuffle () {
-      for (size_t i = length_; i >= 1; --i) {
-         cr::swap (contents_[i - 1], contents_[rg.int_ (i, length_ - 2)]);
+      int32 shuffleLength = length <int32> ();
+
+      for (int32 i = shuffleLength; i >= 1; --i) {
+         cr::swap (contents_[i - 1], contents_[rg.get (i, shuffleLength - 2)]);
       }
    }
 
@@ -352,11 +354,11 @@ public:
    }
 
    const T &random () const {
-      return contents_[rg.int_ <size_t> (0, length_ - 1)];
+      return contents_[rg.get (0, length <int32> () - 1)];
    }
 
    T &random () {
-      return contents_[rg.int_ <size_t> (0u, length_ - 1u)];
+      return contents_[rg.get (0, length <int32> () - 1)];
    }
 
    T *data () {

ext/crlib/cr-http.h

@@ -443,7 +443,7 @@ public:
       if (boundarySlash != String::InvalidIndex) {
          boundaryName = localPath.substr (boundarySlash + 1);
       }
-      StringRef boundaryLine = strings.format ("---crlib_upload_boundary_%d%d%d%d", rg.int_ (0, 9), rg.int_ (0, 9), rg.int_ (0, 9), rg.int_ (0, 9));
+      StringRef boundaryLine = strings.format ("---crlib_upload_boundary_%d%d%d%d", rg.get (0, 9), rg.get (0, 9), rg.get (0, 9), rg.get (0, 9));
 
       String request, start, end;
       start.appendf ("--%s\r\n", boundaryLine);

ext/crlib/cr-random.h

@@ -20,49 +20,57 @@
 
 CR_NAMESPACE_BEGIN
 
-// random number generator see: https://github.com/preshing/RandomSequence/
-class Random final : public Singleton <Random> {
+// based on: https://github.com/jeudesprits/PSWyhash/blob/master/Sources/CWyhash/include/wyhash.h
+class Random : public Singleton <Random> {
 private:
-   uint32 index_, offset_;
-   uint64 divider_;
+   uint64 div_ { static_cast <uint64> (1) << 32ull };
+   uint64 state_ { static_cast <uint64> (time (nullptr)) };
 
 public:
-   explicit Random () {
-      const auto base = static_cast <uint32> (time (nullptr));
-      const auto offset = base + 1;
-
-      index_ = premute (premute (base) + 0x682f0161);
-      offset_ = premute (premute (offset) + 0x46790905);
-      divider_ = (static_cast <uint64> (1)) << 32;
-   }
+   Random () = default;
    ~Random () = default;
 
 private:
-   uint32 premute (uint32 index) {
-      static constexpr auto prime = 4294967291u;
+   uint64 wyrand64 () {
+      constexpr uint64 wyp0 = 0xa0761d6478bd642full, wyp1 = 0xe7037ed1a0b428dbull;
+      state_ += wyp0;
 
-      if (index >= prime) {
-         return index;
-      }
-      const uint32 residue = (static_cast <uint64> (index) * index) % prime;
-      return (index <= prime / 2) ? residue : prime - residue;
+      return mul (state_ ^ wyp1, state_);
    }
 
-   uint32 generate () {
-      return premute ((premute (index_++) + offset_) ^ 0x5bf03635);
+   uint32 wyrand32 () {
+      return static_cast <uint32> (wyrand64 ());
+   }
+
+private:
+   uint64_t rotr (uint64 v, uint32 k) {
+      return (v >> k) | (v << (64 - k));
+   }
+
+   uint64 mul (uint64 a, uint64 b) {
+      uint64 hh = (a >> 32) * (b >> 32);
+      uint64 hl = (b >> 32) * static_cast <uint32> (b);
+
+      uint64 lh = static_cast <uint32> (a) * (b >> 32);
+      uint64 ll = static_cast <uint64> (static_cast <double> (a) * static_cast <double> (b));
+
+      return rotr (hl, 32) ^ rotr (lh, 32) ^ hh ^ ll;
    }
 
 public:
-   template <typename U> U int_ (U low, U high) {
-      return static_cast <U> (generate () * (static_cast <double> (high) - static_cast <double> (low) + 1.0) / divider_ + static_cast <double> (low));
+   template <typename U, typename Void = void> U get (U, U) = delete;
+
+   template <typename Void = void> int32 get (int32 low, int32 high) {
+      return static_cast <int32> (wyrand32 () * (static_cast <double> (high) - static_cast <double> (low) + 1.0) / div_ + static_cast <double> (low));
    }
 
-   float float_ (float low, float high) {
-      return static_cast <float> (generate () * (static_cast <double> (high) - static_cast <double> (low)) / (divider_ - 1) + static_cast <double> (low));
+   template <typename Void = void> float get (float low, float high) {
+      return static_cast <float> (wyrand32 () * (static_cast <double> (high) - static_cast <double> (low)) / (div_ - 1) + static_cast <double> (low));
    }
 
-   template <typename U> bool chance (const U max, const U maxChance = 100) {
-      return int_ <U> (0, maxChance) < max;
+public:
+   bool chance (int32 limit) {
+      return get <int32> (0, 100) < limit;
    }
 };
 

ext/crlib/cr-vector.h

@@ -53,8 +53,6 @@ public:
    SimdWrap (__m128 m) : m (m)
    { }
 
-
-
 public:
    SimdWrap normalize () {
       return { _mm_div_ps (m, _mm_sqrt_ps (wrap_dp_sse2 (m, m))) };