Rewrote the PRNG using an xorshift* algorithm and moved the files from test/ to base/.

webrtc/base/random.h

 /*
  *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef WEBRTC_TEST_RANDOM_H_
-#define WEBRTC_TEST_RANDOM_H_
+#ifndef WEBRTC_BASE_RANDOM_H_
+#define WEBRTC_BASE_RANDOM_H_
 
 #include <limits>
 
 #include "webrtc/typedefs.h"
 #include "webrtc/base/constructormagic.h"
+#include "webrtc/base/checks.h"
 
 namespace webrtc {
 
-namespace test {
-
 class Random {
  public:
-  explicit Random(uint32_t seed);
+  explicit Random(uint64_t seed);
 
   // Return pseudo-random integer of the specified type.
+  // We need to limit the size to 32 bits to keep the output close to uniform.
   template <typename T>
   T Rand() {

[pbos-webrtc, 2015/11/20 12:51:57]

Do we need this generic template version or is support for the specialized
versions enough?

[the sun, 2015/11/25 10:12:37]

It is quite useful; often you need to get a random value for an integer range.

[terelius, 2015/11/27 11:31:23]

I think pbos is suggesting not having the generic rand<T>(), but instead adding
more specialized versions with rand<int>(), rand<unsigned>(), rand<short>(),
etc. The effect would be the same, but the code would be longer.

I prefer rand<T>() unless the specializations somehow solves the
implementation-defined behavior of the cast.

     static_assert(std::numeric_limits<T>::is_integer &&
                       std::numeric_limits<T>::radix == 2 &&
                       std::numeric_limits<T>::digits <= 32,
                   "Rand is only supported for built-in integer types that are "
                   "32 bits or smaller.");
-    return static_cast<T>(Rand(std::numeric_limits<uint32_t>::max()));
+    return static_cast<T>(NextOutput());

[pbos-webrtc, 2015/11/20 12:51:57]

I believe this is undefined value/implementation may raise signal (almost UB)
for uint32_t->int32_t for half the values.

[pbos-webrtc, 2015/11/25 21:16:16]

We might accept this so long as ubsan doesn't start barking on it. It's not
undefined behaviour, but it may raise a trap which would effectively crash the
program. I don't think we run on any such platforms.

[terelius, 2015/11/27 11:31:23]

The best solution I've been able to come up with, is 
uint32_t r = NextOutput();
return *((T*) &r);
which forces a uniform bit pattern into a T. However, trying to read the
resulting T is probably undefined since the T could have padding bits which the
processor/compiler might expect to be e.g. zero. I don't know of any such
machines.

I'm pretty sure we don't run on machines where a plain static_cast from unsigned
to signed causes a trap, since there are numerous of those casts in the code
base already.

   }
 
   // Uniformly distributed pseudo-random number in the interval [0, t].
   uint32_t Rand(uint32_t t);
 
   // Uniformly distributed pseudo-random number in the interval [low, high].
   uint32_t Rand(uint32_t low, uint32_t high);
 
+  // Uniformly distributed pseudo-random number in the interval [low, high].
+  int32_t Rand(int32_t low, int32_t high);
+
   // Normal Distribution.
-  int Gaussian(int mean, int standard_deviation);
+  double Gaussian(double mean, double standard_deviation);
 
   // Exponential Distribution.
-  int Exponential(float lambda);
-
-  // TODO(solenberg): Random from histogram.
-  // template<typename T> int Distribution(const std::vector<T> histogram) {
+  double Exponential(double lambda);
 
  private:
-  uint32_t a_;
-  uint32_t b_;
+  // Outputs a nonzero 64-bit random number.
+  uint64_t NextOutput() {

[pbos-webrtc, 2015/11/20 12:51:57]

Move implementation to .cc

[terelius, 2015/11/27 11:31:23]

I actually prefer to have this inlined since the body should take ~10 cycles.
The cost of an extra function call would be non-negligible I think. 

+    state_ ^= state_ >> 12;
+    state_ ^= state_ << 25;
+    state_ ^= state_ >> 27;
+    RTC_DCHECK(state_ != 0x0ULL);
+    return state_ * 2685821657736338717ull;
+  }
+
+  uint64_t state_;
 
   RTC_DISALLOW_IMPLICIT_CONSTRUCTORS(Random);
 };
 
 // Return pseudo-random number in the interval [0.0, 1.0).
 template <>
 float Random::Rand<float>();
 
+// Return pseudo-random number in the interval [0.0, 1.0).
+template <>
+double Random::Rand<double>();
+
 // Return pseudo-random boolean value.
 template <>
 bool Random::Rand<bool>();
 
-}  // namespace test
 }  // namespace webrtc
 
-#endif  // WEBRTC_TEST_RANDOM_H_
+#endif  // WEBRTC_BASE_RANDOM_H_