diff --git a/absl/base/BUILD.bazel b/absl/base/BUILD.bazel
index d74bac631a6..ffeca630706 100644
--- a/absl/base/BUILD.bazel
+++ b/absl/base/BUILD.bazel
@@ -557,6 +557,31 @@ cc_test(
     ],
 )
 
+cc_library(
+    name = "exponential_biased",
+    srcs = ["internal/exponential_biased.cc"],
+    hdrs = ["internal/exponential_biased.h"],
+    linkopts = ABSL_DEFAULT_LINKOPTS,
+    visibility = [
+        "//absl:__subpackages__",
+    ],
+    deps = [":core_headers"],
+)
+
+cc_test(
+    name = "exponential_biased_test",
+    size = "small",
+    srcs = ["internal/exponential_biased_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    linkopts = ABSL_DEFAULT_LINKOPTS,
+    visibility = ["//visibility:private"],
+    deps = [
+        ":exponential_biased",
+        "//absl/strings",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
 cc_library(
     name = "scoped_set_env",
     testonly = 1,
diff --git a/absl/base/CMakeLists.txt b/absl/base/CMakeLists.txt
index 79ee5b93ed7..2698213c566 100644
--- a/absl/base/CMakeLists.txt
+++ b/absl/base/CMakeLists.txt
@@ -503,6 +503,32 @@ absl_cc_test(
     gtest_main
 )
 
+absl_cc_library(
+  NAME
+    exponential_biased
+  SRCS
+    "internal/exponential_biased.cc"
+  HDRS
+    "internal/exponential_biased.h"
+  COPTS
+    ${ABSL_DEFAULT_COPTS}
+  DEPS
+    absl::core_headers
+)
+
+absl_cc_test(
+  NAME
+    exponential_biased_test
+  SRCS
+    "internal/exponential_biased_test.cc"
+  COPTS
+    ${ABSL_TEST_COPTS}
+  DEPS
+    absl::exponential_biased
+    absl::strings
+    gmock_main
+)
+
 absl_cc_library(
   NAME
     scoped_set_env
diff --git a/absl/base/config.h b/absl/base/config.h
index 039b73d81f7..24851fa33ce 100644
--- a/absl/base/config.h
+++ b/absl/base/config.h
@@ -307,7 +307,7 @@
 
 // ABSL_HAVE_SEMAPHORE_H
 //
-// Checks whether the platform supports the <semaphore.h> header and sem_open(3)
+// Checks whether the platform supports the <semaphore.h> header and sem_init(3)
 // family of functions as standardized in POSIX.1-2001.
 //
 // Note: While Apple provides <semaphore.h> for both iOS and macOS, it is
diff --git a/absl/base/internal/exponential_biased.cc b/absl/base/internal/exponential_biased.cc
new file mode 100644
index 00000000000..d7ffd184e96
--- /dev/null
+++ b/absl/base/internal/exponential_biased.cc
@@ -0,0 +1,84 @@
+// Copyright 2019 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/base/internal/exponential_biased.h"
+
+#include <stdint.h>
+
+#include <atomic>
+#include <cmath>
+#include <limits>
+
+#include "absl/base/attributes.h"
+#include "absl/base/optimization.h"
+
+namespace absl {
+namespace base_internal {
+
+// The algorithm generates a random number between 0 and 1 and applies the
+// inverse cumulative distribution function for an exponential. Specifically:
+// Let m be the inverse of the sample period, then the probability
+// distribution function is m*exp(-mx) so the CDF is
+// p = 1 - exp(-mx), so
+// q = 1 - p = exp(-mx)
+// log_e(q) = -mx
+// -log_e(q)/m = x
+// log_2(q) * (-log_e(2) * 1/m) = x
+// In the code, q is actually in the range 1 to 2**26, hence the -26 below
+int64_t ExponentialBiased::Get(int64_t mean) {
+  if (ABSL_PREDICT_FALSE(!initialized_)) {
+    Initialize();
+  }
+
+  uint64_t rng = NextRandom(rng_);
+  rng_ = rng;
+
+  // Take the top 26 bits as the random number
+  // (This plus the 1<<58 sampling bound give a max possible step of
+  // 5194297183973780480 bytes.)
+  // The uint32_t cast is to prevent a (hard-to-reproduce) NAN
+  // under piii debug for some binaries.
+  double q = static_cast<uint32_t>(rng >> (kPrngNumBits - 26)) + 1.0;
+  // Put the computed p-value through the CDF of a geometric.
+  double interval = (std::log2(q) - 26) * (-std::log(2.0) * mean);
+  // Very large values of interval overflow int64_t. To avoid that, we will cheat
+  // and clamp any huge values to (int64_t max)/2. This is a potential source of
+  // bias, but the mean would need to be such a large value that it's not likely
+  // to come up. For example, with a mean of 1e18, the probability of hitting
+  // this condition is about 1/1000. For a mean of 1e17, standard calculators
+  // claim that this event won't happen.
+  if (interval > static_cast<double>(std::numeric_limits<int64_t>::max() / 2)) {
+    return std::numeric_limits<int64_t>::max() / 2;
+  }
+
+  return static_cast<int64_t>(interval);
+}
+
+void ExponentialBiased::Initialize() {
+  // We don't get well distributed numbers from `this` so we call NextRandom() a
+  // bunch to mush the bits around. We use a global_rand to handle the case
+  // where the same thread (by memory address) gets created and destroyed
+  // repeatedly.
+  ABSL_CONST_INIT static std::atomic<uint32_t> global_rand(0);
+  uint64_t r = reinterpret_cast<uint64_t>(this) +
+               global_rand.fetch_add(1, std::memory_order_relaxed);
+  for (int i = 0; i < 20; ++i) {
+    r = NextRandom(r);
+  }
+  rng_ = r;
+  initialized_ = true;
+}
+
+}  // namespace base_internal
+}  // namespace absl
diff --git a/absl/base/internal/exponential_biased.h b/absl/base/internal/exponential_biased.h
new file mode 100644
index 00000000000..cac2d8ad84f
--- /dev/null
+++ b/absl/base/internal/exponential_biased.h
@@ -0,0 +1,77 @@
+// Copyright 2019 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ABSL_BASE_INTERNAL_EXPONENTIAL_BIASED_H_
+#define ABSL_BASE_INTERNAL_EXPONENTIAL_BIASED_H_
+
+#include <stdint.h>
+
+namespace absl {
+namespace base_internal {
+
+// ExponentialBiased provides a small and fast random number generator for a
+// rounded exponential distribution. This generator doesn't requires very little
+// state doesn't impose synchronization overhead, which makes it useful in some
+// specialized scenarios.
+//
+// For the generated variable X, X ~ floor(Exponential(1/mean)). The floor
+// operation introduces a small amount of bias, but the distribution is useful
+// to generate a wait time. That is, if an operation is supposed to happen on
+// average to 1/mean events, then the generated variable X will describe how
+// many events to skip before performing the operation and computing a new X.
+//
+// The mathematically precise distribution to use for integer wait times is a
+// Geometric distribution, but a Geometric distribution takes slightly more time
+// to compute and when the mean is large (say, 100+), the Geometric distribution
+// is hard to distinguish from the result of ExponentialBiased.
+//
+// This class is thread-compatible.
+class ExponentialBiased {
+ public:
+  // The number of bits set by NextRandom.
+  static constexpr int kPrngNumBits = 48;
+
+  // Generates the floor of an exponentially distributed random variable by
+  // rounding the value down to the nearest integer. The result will be in the
+  // range [0, int64_t max / 2].
+  int64_t Get(int64_t mean);
+
+  // Computes a random number in the range [0, 1<<(kPrngNumBits+1) - 1]
+  //
+  // This is public to enable testing.
+  static uint64_t NextRandom(uint64_t rnd);
+
+ private:
+  void Initialize();
+
+  uint64_t rng_{0};
+  bool initialized_{false};
+};
+
+// Returns the next prng value.
+// pRNG is: aX+b mod c with a = 0x5DEECE66D, b =  0xB, c = 1<<48
+// This is the lrand64 generator.
+inline uint64_t ExponentialBiased::NextRandom(uint64_t rnd) {
+  const uint64_t prng_mult = uint64_t{0x5DEECE66D};
+  const uint64_t prng_add = 0xB;
+  const uint64_t prng_mod_power = 48;
+  const uint64_t prng_mod_mask =
+      ~((~static_cast<uint64_t>(0)) << prng_mod_power);
+  return (prng_mult * rnd + prng_add) & prng_mod_mask;
+}
+
+}  // namespace base_internal
+}  // namespace absl
+
+#endif  // ABSL_BASE_INTERNAL_EXPONENTIAL_BIASED_H_
diff --git a/absl/base/internal/exponential_biased_test.cc b/absl/base/internal/exponential_biased_test.cc
new file mode 100644
index 00000000000..09b511d14e7
--- /dev/null
+++ b/absl/base/internal/exponential_biased_test.cc
@@ -0,0 +1,168 @@
+// Copyright 2019 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/base/internal/exponential_biased.h"
+
+#include <stddef.h>
+
+#include <cmath>
+#include <cstdint>
+#include <vector>
+
+#include "gmock/gmock.h"
+#include "gtest/gtest.h"
+#include "absl/strings/str_cat.h"
+
+using ::testing::Ge;
+
+namespace absl {
+namespace base_internal {
+
+MATCHER_P2(IsBetween, a, b,
+           absl::StrCat(std::string(negation ? "isn't" : "is"), " between ", a,
+                        " and ", b)) {
+  return a <= arg && arg <= b;
+}
+
+// Tests of the quality of the random numbers generated
+// This uses the Anderson Darling test for uniformity.
+// See "Evaluating the Anderson-Darling Distribution" by Marsaglia
+// for details.
+
+// Short cut version of ADinf(z), z>0 (from Marsaglia)
+// This returns the p-value for Anderson Darling statistic in
+// the limit as n-> infinity. For finite n, apply the error fix below.
+double AndersonDarlingInf(double z) {
+  if (z < 2) {
+    return exp(-1.2337141 / z) / sqrt(z) *
+           (2.00012 +
+            (0.247105 -
+             (0.0649821 - (0.0347962 - (0.011672 - 0.00168691 * z) * z) * z) *
+                 z) *
+                z);
+  }
+  return exp(
+      -exp(1.0776 -
+           (2.30695 -
+            (0.43424 - (0.082433 - (0.008056 - 0.0003146 * z) * z) * z) * z) *
+               z));
+}
+
+// Corrects the approximation error in AndersonDarlingInf for small values of n
+// Add this to AndersonDarlingInf to get a better approximation
+// (from Marsaglia)
+double AndersonDarlingErrFix(int n, double x) {
+  if (x > 0.8) {
+    return (-130.2137 +
+            (745.2337 -
+             (1705.091 - (1950.646 - (1116.360 - 255.7844 * x) * x) * x) * x) *
+                x) /
+           n;
+  }
+  double cutoff = 0.01265 + 0.1757 / n;
+  if (x < cutoff) {
+    double t = x / cutoff;
+    t = sqrt(t) * (1 - t) * (49 * t - 102);
+    return t * (0.0037 / (n * n) + 0.00078 / n + 0.00006) / n;
+  } else {
+    double t = (x - cutoff) / (0.8 - cutoff);
+    t = -0.00022633 +
+        (6.54034 - (14.6538 - (14.458 - (8.259 - 1.91864 * t) * t) * t) * t) *
+            t;
+    return t * (0.04213 + 0.01365 / n) / n;
+  }
+}
+
+// Returns the AndersonDarling p-value given n and the value of the statistic
+double AndersonDarlingPValue(int n, double z) {
+  double ad = AndersonDarlingInf(z);
+  double errfix = AndersonDarlingErrFix(n, ad);
+  return ad + errfix;
+}
+
+double AndersonDarlingStatistic(const std::vector<double>& random_sample) {
+  int n = random_sample.size();
+  double ad_sum = 0;
+  for (int i = 0; i < n; i++) {
+    ad_sum += (2 * i + 1) *
+              std::log(random_sample[i] * (1 - random_sample[n - 1 - i]));
+  }
+  double ad_statistic = -n - 1 / static_cast<double>(n) * ad_sum;
+  return ad_statistic;
+}
+
+// Tests if the array of doubles is uniformly distributed.
+// Returns the p-value of the Anderson Darling Statistic
+// for the given set of sorted random doubles
+// See "Evaluating the Anderson-Darling Distribution" by
+// Marsaglia and Marsaglia for details.
+double AndersonDarlingTest(const std::vector<double>& random_sample) {
+  double ad_statistic = AndersonDarlingStatistic(random_sample);
+  double p = AndersonDarlingPValue(random_sample.size(), ad_statistic);
+  return p;
+}
+
+// Testing that NextRandom generates uniform random numbers. Applies the
+// Anderson-Darling test for uniformity
+TEST(ExponentialBiasedTest, TestNextRandom) {
+  for (auto n : std::vector<int>({
+           10,  // Check short-range correlation
+           100, 1000,
+           10000  // Make sure there's no systemic error
+       })) {
+    uint64_t x = 1;
+    // This assumes that the prng returns 48 bit numbers
+    uint64_t max_prng_value = static_cast<uint64_t>(1) << 48;
+    // Initialize.
+    for (int i = 1; i <= 20; i++) {
+      x = ExponentialBiased::NextRandom(x);
+    }
+    std::vector<uint64_t> int_random_sample(n);
+    // Collect samples
+    for (int i = 0; i < n; i++) {
+      int_random_sample[i] = x;
+      x = ExponentialBiased::NextRandom(x);
+    }
+    // First sort them...
+    std::sort(int_random_sample.begin(), int_random_sample.end());
+    std::vector<double> random_sample(n);
+    // Convert them to uniform randoms (in the range [0,1])
+    for (int i = 0; i < n; i++) {
+      random_sample[i] =
+          static_cast<double>(int_random_sample[i]) / max_prng_value;
+    }
+    // Now compute the Anderson-Darling statistic
+    double ad_pvalue = AndersonDarlingTest(random_sample);
+    EXPECT_GT(std::min(ad_pvalue, 1 - ad_pvalue), 0.0001)
+        << "prng is not uniform: n = " << n << " p = " << ad_pvalue;
+  }
+}
+
+// The generator needs to be available as a thread_local and as a static
+// variable.
+TEST(ExponentialBiasedTest, InitializationModes) {
+  ABSL_CONST_INIT static ExponentialBiased eb_static;
+  EXPECT_THAT(eb_static.Get(2), Ge(0));
+
+#if ABSL_HAVE_THREAD_LOCAL
+  thread_local ExponentialBiased eb_thread;
+  EXPECT_THAT(eb_thread.Get(2), Ge(0));
+#endif
+
+  ExponentialBiased eb_stack;
+  EXPECT_THAT(eb_stack.Get(2), Ge(0));
+}
+
+}  // namespace base_internal
+}  // namespace absl
diff --git a/absl/base/optimization.h b/absl/base/optimization.h
index 0dcbef32a3c..646523b3461 100644
--- a/absl/base/optimization.h
+++ b/absl/base/optimization.h
@@ -172,7 +172,7 @@
 #if ABSL_HAVE_BUILTIN(__builtin_expect) || \
     (defined(__GNUC__) && !defined(__clang__))
 #define ABSL_PREDICT_FALSE(x) (__builtin_expect(x, 0))
-#define ABSL_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
+#define ABSL_PREDICT_TRUE(x) (__builtin_expect(false || (x), true))
 #else
 #define ABSL_PREDICT_FALSE(x) (x)
 #define ABSL_PREDICT_TRUE(x) (x)
diff --git a/absl/container/BUILD.bazel b/absl/container/BUILD.bazel
index 0894bb207c6..e60979b2a54 100644
--- a/absl/container/BUILD.bazel
+++ b/absl/container/BUILD.bazel
@@ -493,6 +493,7 @@ cc_library(
         ":have_sse",
         "//absl/base",
         "//absl/base:core_headers",
+        "//absl/base:exponential_biased",
         "//absl/debugging:stacktrace",
         "//absl/memory",
         "//absl/synchronization",
diff --git a/absl/container/CMakeLists.txt b/absl/container/CMakeLists.txt
index 933c7a8c48a..aa33659b322 100644
--- a/absl/container/CMakeLists.txt
+++ b/absl/container/CMakeLists.txt
@@ -538,6 +538,7 @@ absl_cc_library(
     ${ABSL_DEFAULT_COPTS}
   DEPS
     absl::base
+    absl::exponential_biased
     absl::have_sse
     absl::synchronization
 )
diff --git a/absl/container/flat_hash_map.h b/absl/container/flat_hash_map.h
index 5c16ac88fbb..283f2439676 100644
--- a/absl/container/flat_hash_map.h
+++ b/absl/container/flat_hash_map.h
@@ -401,7 +401,7 @@ class flat_hash_map : public absl::container_internal::raw_hash_map<
   // for the past-the-end iterator, which is invalidated.
   //
   // `swap()` requires that the flat hash map's hashing and key equivalence
-  // functions be Swappable, and are exchaged using unqualified calls to
+  // functions be Swappable, and are exchanged using unqualified calls to
   // non-member `swap()`. If the map's allocator has
   // `std::allocator_traits<allocator_type>::propagate_on_container_swap::value`
   // set to `true`, the allocators are also exchanged using an unqualified call
diff --git a/absl/container/internal/hashtablez_sampler.cc b/absl/container/internal/hashtablez_sampler.cc
index 054e8981781..0a7ef61c8fe 100644
--- a/absl/container/internal/hashtablez_sampler.cc
+++ b/absl/container/internal/hashtablez_sampler.cc
@@ -21,6 +21,7 @@
 #include <limits>
 
 #include "absl/base/attributes.h"
+#include "absl/base/internal/exponential_biased.h"
 #include "absl/container/internal/have_sse.h"
 #include "absl/debugging/stacktrace.h"
 #include "absl/memory/memory.h"
@@ -37,77 +38,13 @@ ABSL_CONST_INIT std::atomic<bool> g_hashtablez_enabled{
 ABSL_CONST_INIT std::atomic<int32_t> g_hashtablez_sample_parameter{1 << 10};
 ABSL_CONST_INIT std::atomic<int32_t> g_hashtablez_max_samples{1 << 20};
 
-// Returns the next pseudo-random value.
-// pRNG is: aX+b mod c with a = 0x5DEECE66D, b =  0xB, c = 1<<48
-// This is the lrand64 generator.
-uint64_t NextRandom(uint64_t rnd) {
-  const uint64_t prng_mult = uint64_t{0x5DEECE66D};
-  const uint64_t prng_add = 0xB;
-  const uint64_t prng_mod_power = 48;
-  const uint64_t prng_mod_mask = ~(~uint64_t{0} << prng_mod_power);
-  return (prng_mult * rnd + prng_add) & prng_mod_mask;
-}
-
-// Generates a geometric variable with the specified mean.
-// This is done by generating a random number between 0 and 1 and applying
-// the inverse cumulative distribution function for an exponential.
-// Specifically: Let m be the inverse of the sample period, then
-// the probability distribution function is m*exp(-mx) so the CDF is
-// p = 1 - exp(-mx), so
-// q = 1 - p = exp(-mx)
-// log_e(q) = -mx
-// -log_e(q)/m = x
-// log_2(q) * (-log_e(2) * 1/m) = x
-// In the code, q is actually in the range 1 to 2**26, hence the -26 below
-//
-int64_t GetGeometricVariable(int64_t mean) {
 #if ABSL_HAVE_THREAD_LOCAL
-  thread_local
-#else   // ABSL_HAVE_THREAD_LOCAL
-  // SampleSlow and hence GetGeometricVariable is guarded by a single mutex when
-  // there are not thread locals.  Thus, a single global rng is acceptable for
-  // that case.
-  static
-#endif  // ABSL_HAVE_THREAD_LOCAL
-      uint64_t rng = []() {
-        // We don't get well distributed numbers from this so we call
-        // NextRandom() a bunch to mush the bits around.  We use a global_rand
-        // to handle the case where the same thread (by memory address) gets
-        // created and destroyed repeatedly.
-        ABSL_CONST_INIT static std::atomic<uint32_t> global_rand(0);
-        uint64_t r = reinterpret_cast<uint64_t>(&rng) +
-                   global_rand.fetch_add(1, std::memory_order_relaxed);
-        for (int i = 0; i < 20; ++i) {
-          r = NextRandom(r);
-        }
-        return r;
-      }();
-
-  rng = NextRandom(rng);
-
-  // Take the top 26 bits as the random number
-  // (This plus the 1<<58 sampling bound give a max possible step of
-  // 5194297183973780480 bytes.)
-  const uint64_t prng_mod_power = 48;  // Number of bits in prng
-  // The uint32_t cast is to prevent a (hard-to-reproduce) NAN
-  // under piii debug for some binaries.
-  double q = static_cast<uint32_t>(rng >> (prng_mod_power - 26)) + 1.0;
-  // Put the computed p-value through the CDF of a geometric.
-  double interval = (log2(q) - 26) * (-std::log(2.0) * mean);
-
-  // Very large values of interval overflow int64_t. If we happen to
-  // hit such improbable condition, we simply cheat and clamp interval
-  // to largest supported value.
-  if (interval > static_cast<double>(std::numeric_limits<int64_t>::max() / 2)) {
-    return std::numeric_limits<int64_t>::max() / 2;
-  }
-
-  // Small values of interval are equivalent to just sampling next time.
-  if (interval < 1) {
-    return 1;
-  }
-  return static_cast<int64_t>(interval);
-}
+thread_local absl::base_internal::ExponentialBiased
+    g_exponential_biased_generator;
+#else
+ABSL_CONST_INIT static absl::base_internal::ExponentialBiased
+    g_exponential_biased_generator;
+#endif
 
 }  // namespace
 
@@ -253,8 +190,12 @@ HashtablezInfo* SampleSlow(int64_t* next_sample) {
   }
 
   bool first = *next_sample < 0;
-  *next_sample = GetGeometricVariable(
+  *next_sample = g_exponential_biased_generator.Get(
       g_hashtablez_sample_parameter.load(std::memory_order_relaxed));
+  // Small values of interval are equivalent to just sampling next time.
+  if (*next_sample < 1) {
+    *next_sample = 1;
+  }
 
   // g_hashtablez_enabled can be dynamically flipped, we need to set a threshold
   // low enough that we will start sampling in a reasonable time, so we just use
diff --git a/absl/container/internal/raw_hash_set.h b/absl/container/internal/raw_hash_set.h
index 42b3c468113..9992ba4b8ac 100644
--- a/absl/container/internal/raw_hash_set.h
+++ b/absl/container/internal/raw_hash_set.h
@@ -614,13 +614,17 @@ class raw_hash_set {
     iterator() {}
 
     // PRECONDITION: not an end() iterator.
-    reference operator*() const { return PolicyTraits::element(slot_); }
+    reference operator*() const {
+      /* To be enabled: assert_is_full(); */
+      return PolicyTraits::element(slot_);
+    }
 
     // PRECONDITION: not an end() iterator.
     pointer operator->() const { return &operator*(); }
 
     // PRECONDITION: not an end() iterator.
     iterator& operator++() {
+      /* To be enabled: assert_is_full(); */
       ++ctrl_;
       ++slot_;
       skip_empty_or_deleted();
@@ -634,6 +638,8 @@ class raw_hash_set {
     }
 
     friend bool operator==(const iterator& a, const iterator& b) {
+      /* To be enabled: a.assert_is_valid(); */
+      /* To be enabled: b.assert_is_valid(); */
       return a.ctrl_ == b.ctrl_;
     }
     friend bool operator!=(const iterator& a, const iterator& b) {
@@ -644,6 +650,11 @@ class raw_hash_set {
     iterator(ctrl_t* ctrl) : ctrl_(ctrl) {}  // for end()
     iterator(ctrl_t* ctrl, slot_type* slot) : ctrl_(ctrl), slot_(slot) {}
 
+    void assert_is_full() const { assert(IsFull(*ctrl_)); }
+    void assert_is_valid() const {
+      assert(!ctrl_ || IsFull(*ctrl_) || *ctrl_ == kSentinel);
+    }
+
     void skip_empty_or_deleted() {
       while (IsEmptyOrDeleted(*ctrl_)) {
         // ctrl is not necessarily aligned to Group::kWidth. It is also likely
@@ -1155,7 +1166,7 @@ class raw_hash_set {
   // This overload is necessary because otherwise erase<K>(const K&) would be
   // a better match if non-const iterator is passed as an argument.
   void erase(iterator it) {
-    assert(it != end());
+    it.assert_is_full();
     PolicyTraits::destroy(&alloc_ref(), it.slot_);
     erase_meta_only(it);
   }
@@ -1172,12 +1183,14 @@ class raw_hash_set {
   template <typename H, typename E>
   void merge(raw_hash_set<Policy, H, E, Alloc>& src) {  // NOLINT
     assert(this != &src);
-    for (auto it = src.begin(), e = src.end(); it != e; ++it) {
+    for (auto it = src.begin(), e = src.end(); it != e;) {
+      auto next = std::next(it);
       if (PolicyTraits::apply(InsertSlot<false>{*this, std::move(*it.slot_)},
                               PolicyTraits::element(it.slot_))
               .second) {
         src.erase_meta_only(it);
       }
+      it = next;
     }
   }
 
@@ -1187,6 +1200,7 @@ class raw_hash_set {
   }
 
   node_type extract(const_iterator position) {
+    position.inner_.assert_is_full();
     auto node =
         CommonAccess::Transfer<node_type>(alloc_ref(), position.inner_.slot_);
     erase_meta_only(position);
diff --git a/absl/container/internal/raw_hash_set_test.cc b/absl/container/internal/raw_hash_set_test.cc
index ed4ca8c88d3..33cfa72ce7a 100644
--- a/absl/container/internal/raw_hash_set_test.cc
+++ b/absl/container/internal/raw_hash_set_test.cc
@@ -1837,7 +1837,7 @@ TEST(TableDeathTest, EraseOfEndAsserts) {
 
   IntTable t;
   // Extra simple "regexp" as regexp support is highly varied across platforms.
-  constexpr char kDeathMsg[] = "it != end";
+  constexpr char kDeathMsg[] = "IsFull";
   EXPECT_DEATH_IF_SUPPORTED(t.erase(t.end()), kDeathMsg);
 }
 
diff --git a/absl/copts/AbseilConfigureCopts.cmake b/absl/copts/AbseilConfigureCopts.cmake
index b430873c5e0..b442646948a 100644
--- a/absl/copts/AbseilConfigureCopts.cmake
+++ b/absl/copts/AbseilConfigureCopts.cmake
@@ -5,10 +5,29 @@ set(ABSL_LSAN_LINKOPTS "")
 set(ABSL_HAVE_LSAN OFF)
 set(ABSL_DEFAULT_LINKOPTS "")
 
+if("${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "x86_64")
+  if (MSVC)
+    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_MSVC_X64_FLAGS}")
+  else()
+    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_X64_FLAGS}")
+  endif()
+elseif("${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "arm")
+  if ("${CMAKE_SIZEOF_VOID_P}" STREQUAL "8")
+    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_ARM64_FLAGS}")
+  elseif("${CMAKE_SIZEOF_VOID_P}" STREQUAL "4")
+    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_ARM32_FLAGS}")
+  else()
+    message(WARNING "Value of CMAKE_SIZEOF_VOID_P (${CMAKE_SIZEOF_VOID_P}) is not supported.")
+  endif()
+else()
+  message(WARNING "Value of CMAKE_SYSTEM_PROCESSOR (${CMAKE_SYSTEM_PROCESSOR}) is unknown and cannot be used to set ABSL_RANDOM_RANDEN_COPTS")
+  set(ABSL_RANDOM_RANDEN_COPTS "")
+endif()
+
+
 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
   set(ABSL_DEFAULT_COPTS "${ABSL_GCC_FLAGS}")
   set(ABSL_TEST_COPTS "${ABSL_GCC_FLAGS};${ABSL_GCC_TEST_FLAGS}")
-  set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_X64_FLAGS}")
 elseif("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
   # MATCHES so we get both Clang and AppleClang
   if(MSVC)
@@ -16,11 +35,9 @@ elseif("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
     set(ABSL_DEFAULT_COPTS "${ABSL_CLANG_CL_FLAGS}")
     set(ABSL_TEST_COPTS "${ABSL_CLANG_CL_FLAGS};${ABSL_CLANG_CL_TEST_FLAGS}")
     set(ABSL_DEFAULT_LINKOPTS "${ABSL_MSVC_LINKOPTS}")
-    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_MSVC_X64_FLAGS}")
   else()
     set(ABSL_DEFAULT_COPTS "${ABSL_LLVM_FLAGS}")
     set(ABSL_TEST_COPTS "${ABSL_LLVM_FLAGS};${ABSL_LLVM_TEST_FLAGS}")
-    set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_X64_FLAGS}")
     if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
       # AppleClang doesn't have lsan
       # https://developer.apple.com/documentation/code_diagnostics
@@ -34,12 +51,10 @@ elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
   set(ABSL_DEFAULT_COPTS "${ABSL_MSVC_FLAGS}")
   set(ABSL_TEST_COPTS "${ABSL_MSVC_FLAGS};${ABSL_MSVC_TEST_FLAGS}")
   set(ABSL_DEFAULT_LINKOPTS "${ABSL_MSVC_LINKOPTS}")
-  set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_MSVC_X64_FLAGS}")
 else()
   message(WARNING "Unknown compiler: ${CMAKE_CXX_COMPILER}.  Building with no default flags")
   set(ABSL_DEFAULT_COPTS "")
   set(ABSL_TEST_COPTS "")
-  set(ABSL_RANDOM_RANDEN_COPTS "")
 endif()
 
 if("${CMAKE_CXX_STANDARD}" EQUAL 98)
diff --git a/absl/random/BUILD.bazel b/absl/random/BUILD.bazel
index 828f1348bac..c3520df8dcc 100644
--- a/absl/random/BUILD.bazel
+++ b/absl/random/BUILD.bazel
@@ -69,10 +69,10 @@ cc_library(
         "//absl/base:base_internal",
         "//absl/base:core_headers",
         "//absl/meta:type_traits",
-        "//absl/random/internal:distribution_impl",
         "//absl/random/internal:distributions",
         "//absl/random/internal:fast_uniform_bits",
         "//absl/random/internal:fastmath",
+        "//absl/random/internal:generate_real",
         "//absl/random/internal:iostream_state_saver",
         "//absl/random/internal:traits",
         "//absl/random/internal:uniform_helper",
diff --git a/absl/random/CMakeLists.txt b/absl/random/CMakeLists.txt
index 19dd2cabf57..289854ffd40 100644
--- a/absl/random/CMakeLists.txt
+++ b/absl/random/CMakeLists.txt
@@ -58,7 +58,7 @@ absl_cc_library(
   DEPS
     absl::base_internal
     absl::core_headers
-    absl::random_internal_distribution_impl
+    absl::random_internal_generate_real
     absl::random_internal_distributions
     absl::random_internal_fast_uniform_bits
     absl::random_internal_fastmath
@@ -543,19 +543,18 @@ absl_cc_library(
 # Internal-only target, do not depend on directly.
 absl_cc_library(
   NAME
-    random_internal_distribution_impl
+    random_internal_generate_real
   HDRS
-    "internal/distribution_impl.h"
+    "internal/generate_real.h"
   COPTS
     ${ABSL_DEFAULT_COPTS}
   LINKOPTS
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
     absl::bits
-    absl::config
-    absl::int128
     absl::random_internal_fastmath
     absl::random_internal_traits
+    absl::type_traits
 )
 
 # Internal-only target, do not depend on directly.
@@ -767,9 +766,9 @@ absl_cc_test(
 # Internal-only target, do not depend on directly.
 absl_cc_test(
   NAME
-    random_internal_distribution_impl_test
+    random_internal_generate_real_test
   SRCS
-    "internal/distribution_impl_test.cc"
+    "internal/generate_real_test.cc"
   COPTS
     ${ABSL_TEST_COPTS}
   LINKOPTS
@@ -777,8 +776,7 @@ absl_cc_test(
   DEPS
     absl::bits
     absl::flags
-    absl::int128
-    absl::random_internal_distribution_impl
+    absl::random_internal_generate_real
     gtest_main
 )
 
@@ -1029,7 +1027,6 @@ absl_cc_library(
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
     absl::core_headers
-    absl::random_internal_distribution_impl
     absl::random_internal_fast_uniform_bits
     absl::random_internal_iostream_state_saver
     absl::random_internal_traits
diff --git a/absl/random/beta_distribution.h b/absl/random/beta_distribution.h
index e29894f2037..b09b02f0dde 100644
--- a/absl/random/beta_distribution.h
+++ b/absl/random/beta_distribution.h
@@ -22,9 +22,10 @@
 #include <ostream>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
+#include "absl/meta/type_traits.h"
 #include "absl/random/internal/fast_uniform_bits.h"
 #include "absl/random/internal/fastmath.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 
 namespace absl {
@@ -275,15 +276,21 @@ typename beta_distribution<RealType>::result_type
 beta_distribution<RealType>::AlgorithmJoehnk(
     URBG& g,  // NOLINT(runtime/references)
     const param_type& p) {
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+  using real_type =
+      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
+
   // Based on Joehnk, M. D. Erzeugung von betaverteilten und gammaverteilten
   // Zufallszahlen. Metrika 8.1 (1964): 5-15.
   // This method is described in Knuth, Vol 2 (Third Edition), pp 134.
-  using RandU64ToReal = typename random_internal::RandU64ToReal<result_type>;
-  using random_internal::PositiveValueT;
+
   result_type u, v, x, y, z;
   for (;;) {
-    u = RandU64ToReal::template Value<PositiveValueT, false>(fast_u64_(g));
-    v = RandU64ToReal::template Value<PositiveValueT, false>(fast_u64_(g));
+    u = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
+        fast_u64_(g));
+    v = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
+        fast_u64_(g));
 
     // Direct method. std::pow is slow for float, so rely on the optimizer to
     // remove the std::pow() path for that case.
@@ -327,12 +334,14 @@ typename beta_distribution<RealType>::result_type
 beta_distribution<RealType>::AlgorithmCheng(
     URBG& g,  // NOLINT(runtime/references)
     const param_type& p) {
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+  using real_type =
+      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
+
   // Based on Cheng, Russell CH. Generating beta variates with nonintegral
   // shape parameters. Communications of the ACM 21.4 (1978): 317-322.
   // (https://dl.acm.org/citation.cfm?id=359482).
-  using RandU64ToReal = typename random_internal::RandU64ToReal<result_type>;
-  using random_internal::PositiveValueT;
-
   static constexpr result_type kLogFour =
       result_type(1.3862943611198906188344642429163531361);  // log(4)
   static constexpr result_type kS =
@@ -341,8 +350,10 @@ beta_distribution<RealType>::AlgorithmCheng(
   const bool use_algorithm_ba = (p.method_ == param_type::CHENG_BA);
   result_type u1, u2, v, w, z, r, s, t, bw_inv, lhs;
   for (;;) {
-    u1 = RandU64ToReal::template Value<PositiveValueT, false>(fast_u64_(g));
-    u2 = RandU64ToReal::template Value<PositiveValueT, false>(fast_u64_(g));
+    u1 = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
+        fast_u64_(g));
+    u2 = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
+        fast_u64_(g));
     v = p.y_ * std::log(u1 / (1 - u1));
     w = p.a_ * std::exp(v);
     bw_inv = result_type(1) / (p.b_ + w);
diff --git a/absl/random/beta_distribution_test.cc b/absl/random/beta_distribution_test.cc
index 966ad08b5c9..d0111b3e3a8 100644
--- a/absl/random/beta_distribution_test.cc
+++ b/absl/random/beta_distribution_test.cc
@@ -92,7 +92,7 @@ TYPED_TEST(BetaDistributionInterfaceTest, SerializeTest) {
   for (TypeParam alpha : kValues) {
     for (TypeParam beta : kValues) {
       ABSL_INTERNAL_LOG(
-          INFO, absl::StrFormat("Smoke test for Beta(%f, %f)", alpha, beta));
+          INFO, absl::StrFormat("Smoke test for Beta(%a, %a)", alpha, beta));
 
       param_type param(alpha, beta);
       absl::beta_distribution<TypeParam> before(alpha, beta);
diff --git a/absl/random/exponential_distribution.h b/absl/random/exponential_distribution.h
index c8af197575c..24abf57e99d 100644
--- a/absl/random/exponential_distribution.h
+++ b/absl/random/exponential_distribution.h
@@ -21,8 +21,9 @@
 #include <limits>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
+#include "absl/meta/type_traits.h"
 #include "absl/random/internal/fast_uniform_bits.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 
 namespace absl {
@@ -118,9 +119,14 @@ typename exponential_distribution<RealType>::result_type
 exponential_distribution<RealType>::operator()(
     URBG& g,  // NOLINT(runtime/references)
     const param_type& p) {
-  using random_internal::NegativeValueT;
-  const result_type u = random_internal::RandU64ToReal<
-      result_type>::template Value<NegativeValueT, false>(fast_u64_(g));
+  using random_internal::GenerateNegativeTag;
+  using random_internal::GenerateRealFromBits;
+  using real_type =
+      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
+
+  const result_type u = GenerateRealFromBits<real_type, GenerateNegativeTag,
+                                             false>(fast_u64_(g));  // U(-1, 0)
+
   // log1p(-x) is mathematically equivalent to log(1 - x) but has more
   // accuracy for x near zero.
   return p.neg_inv_lambda_ * std::log1p(u);
diff --git a/absl/random/gaussian_distribution.h b/absl/random/gaussian_distribution.h
index 1d1347bce0a..c299e9441c4 100644
--- a/absl/random/gaussian_distribution.h
+++ b/absl/random/gaussian_distribution.h
@@ -28,8 +28,8 @@
 #include <limits>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
 #include "absl/random/internal/fast_uniform_bits.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 
 namespace absl {
@@ -207,12 +207,18 @@ namespace random_internal {
 
 template <typename URBG>
 inline double gaussian_distribution_base::zignor_fallback(URBG& g, bool neg) {
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+
   // This fallback path happens approximately 0.05% of the time.
   double x, y;
   do {
     // kRInv = 1/r, U(0, 1)
-    x = kRInv * std::log(RandU64ToDouble<PositiveValueT, false>(fast_u64_(g)));
-    y = -std::log(RandU64ToDouble<PositiveValueT, false>(fast_u64_(g)));
+    x = kRInv *
+        std::log(GenerateRealFromBits<double, GeneratePositiveTag, false>(
+            fast_u64_(g)));
+    y = -std::log(
+        GenerateRealFromBits<double, GeneratePositiveTag, false>(fast_u64_(g)));
   } while ((y + y) < (x * x));
   return neg ? (x - kR) : (kR - x);
 }
@@ -220,6 +226,10 @@ inline double gaussian_distribution_base::zignor_fallback(URBG& g, bool neg) {
 template <typename URBG>
 inline double gaussian_distribution_base::zignor(
     URBG& g) {  // NOLINT(runtime/references)
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+  using random_internal::GenerateSignedTag;
+
   while (true) {
     // We use a single uint64_t to generate both a double and a strip.
     // These bits are unused when the generated double is > 1/2^5.
@@ -227,7 +237,8 @@ inline double gaussian_distribution_base::zignor(
     // values (those smaller than 1/2^5, which all end up on the left tail).
     uint64_t bits = fast_u64_(g);
     int i = static_cast<int>(bits & kMask);  // pick a random strip
-    double j = RandU64ToDouble<SignedValueT, false>(bits);  // U(-1, 1)
+    double j = GenerateRealFromBits<double, GenerateSignedTag, false>(
+        bits);  // U(-1, 1)
     const double x = j * zg_.x[i];
 
     // Retangular box. Handles >97% of all cases.
@@ -244,7 +255,8 @@ inline double gaussian_distribution_base::zignor(
     }
 
     // i > 0: Wedge samples using precomputed values.
-    double v = RandU64ToDouble<PositiveValueT, false>(fast_u64_(g));  // U(0, 1)
+    double v = GenerateRealFromBits<double, GeneratePositiveTag, false>(
+        fast_u64_(g));  // U(0, 1)
     if ((zg_.f[i + 1] + v * (zg_.f[i] - zg_.f[i + 1])) <
         std::exp(-0.5 * x * x)) {
       return x;
diff --git a/absl/random/internal/BUILD.bazel b/absl/random/internal/BUILD.bazel
index ec58cecd43e..91388d19f3c 100644
--- a/absl/random/internal/BUILD.bazel
+++ b/absl/random/internal/BUILD.bazel
@@ -175,9 +175,9 @@ cc_library(
 )
 
 cc_library(
-    name = "distribution_impl",
+    name = "generate_real",
     hdrs = [
-        "distribution_impl.h",
+        "generate_real.h",
     ],
     copts = ABSL_DEFAULT_COPTS,
     linkopts = ABSL_DEFAULT_LINKOPTS,
@@ -185,8 +185,7 @@ cc_library(
         ":fastmath",
         ":traits",
         "//absl/base:bits",
-        "//absl/base:config",
-        "//absl/numeric:int128",
+        "//absl/meta:type_traits",
     ],
 )
 
@@ -398,16 +397,17 @@ cc_test(
 )
 
 cc_test(
-    name = "distribution_impl_test",
+    name = "generate_real_test",
     size = "small",
-    srcs = ["distribution_impl_test.cc"],
+    srcs = [
+        "generate_real_test.cc",
+    ],
     copts = ABSL_TEST_COPTS,
     linkopts = ABSL_DEFAULT_LINKOPTS,
     deps = [
-        ":distribution_impl",
+        ":generate_real",
         "//absl/base:bits",
         "//absl/flags:flag",
-        "//absl/numeric:int128",
         "@com_google_googletest//:gtest_main",
     ],
 )
diff --git a/absl/random/internal/distribution_impl.h b/absl/random/internal/distribution_impl.h
deleted file mode 100644
index 49b3e1a6e33..00000000000
--- a/absl/random/internal/distribution_impl.h
+++ /dev/null
@@ -1,194 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_
-#define ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_
-
-// This file contains some implementation details which are used by one or more
-// of the absl random number distributions.
-
-#include <cfloat>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <limits>
-#include <type_traits>
-
-#if (defined(_WIN32) || defined(_WIN64)) && defined(_M_IA64)
-#include <intrin.h>  // NOLINT(build/include_order)
-#pragma intrinsic(_umul128)
-#define ABSL_INTERNAL_USE_UMUL128 1
-#endif
-
-#include "absl/base/config.h"
-#include "absl/base/internal/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-namespace random_internal {
-
-// Creates a double from `bits`, with the template fields controlling the
-// output.
-//
-// RandU64To is both more efficient and generates more unique values in the
-// result interval than known implementations of std::generate_canonical().
-//
-// The `Signed` parameter controls whether positive, negative, or both are
-// returned (thus affecting the output interval).
-//   When Signed == SignedValueT, range is U(-1, 1)
-//   When Signed == NegativeValueT, range is U(-1, 0)
-//   When Signed == PositiveValueT, range is U(0, 1)
-//
-// When the `IncludeZero` parameter is true, the function may return 0 for some
-// inputs, otherwise it never returns 0.
-//
-// The `ExponentBias` parameter determines the scale of the output range by
-// adjusting the exponent.
-//
-// When a value in U(0,1) is required, use:
-//   RandU64ToDouble<PositiveValueT, true, 0>();
-//
-// When a value in U(-1,1) is required, use:
-//   RandU64ToDouble<SignedValueT, false, 0>() => U(-1, 1)
-// This generates more distinct values than the mathematically equivalent
-// expression `U(0, 1) * 2.0 - 1.0`, and is preferable.
-//
-// Scaling the result by powers of 2 (and avoiding a multiply) is also possible:
-//   RandU64ToDouble<PositiveValueT, false, 1>();  => U(0, 2)
-//   RandU64ToDouble<PositiveValueT, false, -1>();  => U(0, 0.5)
-//
-
-// Tristate types controlling the output.
-struct PositiveValueT {};
-struct NegativeValueT {};
-struct SignedValueT {};
-
-// RandU64ToDouble is the double-result variant of RandU64To, described above.
-template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-inline double RandU64ToDouble(uint64_t bits) {
-  static_assert(std::is_same<Signed, PositiveValueT>::value ||
-                    std::is_same<Signed, NegativeValueT>::value ||
-                    std::is_same<Signed, SignedValueT>::value,
-                "");
-
-  // Maybe use the left-most bit for a sign bit.
-  uint64_t sign = std::is_same<Signed, NegativeValueT>::value
-                      ? 0x8000000000000000ull
-                      : 0;  // Sign bits.
-
-  if (std::is_same<Signed, SignedValueT>::value) {
-    sign = bits & 0x8000000000000000ull;
-    bits = bits & 0x7FFFFFFFFFFFFFFFull;
-  }
-  if (IncludeZero) {
-    if (bits == 0u) return 0;
-  }
-
-  // Number of leading zeros is mapped to the exponent: 2^-clz
-  int clz = base_internal::CountLeadingZeros64(bits);
-  // Shift number left to erase leading zeros.
-  bits <<= IncludeZero ? clz : (clz & 63);
-
-  // Shift number right to remove bits that overflow double mantissa.  The
-  // direction of the shift depends on `clz`.
-  bits >>= (64 - DBL_MANT_DIG);
-
-  // Compute IEEE 754 double exponent.
-  // In the Signed case, bits is a 63-bit number with a 0 msb.  Adjust the
-  // exponent to account for that.
-  const uint64_t exp =
-      (std::is_same<Signed, SignedValueT>::value ? 1023U : 1022U) +
-      static_cast<uint64_t>(ExponentBias - clz);
-  constexpr int kExp = DBL_MANT_DIG - 1;
-  // Construct IEEE 754 double from exponent and mantissa.
-  const uint64_t val = sign | (exp << kExp) | (bits & ((1ULL << kExp) - 1U));
-
-  double res;
-  static_assert(sizeof(res) == sizeof(val), "double is not 64 bit");
-  // Memcpy value from "val" to "res" to avoid aliasing problems.  Assumes that
-  // endian-ness is same for double and uint64_t.
-  std::memcpy(&res, &val, sizeof(res));
-
-  return res;
-}
-
-// RandU64ToFloat is the float-result variant of RandU64To, described above.
-template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-inline float RandU64ToFloat(uint64_t bits) {
-  static_assert(std::is_same<Signed, PositiveValueT>::value ||
-                    std::is_same<Signed, NegativeValueT>::value ||
-                    std::is_same<Signed, SignedValueT>::value,
-                "");
-
-  // Maybe use the left-most bit for a sign bit.
-  uint64_t sign = std::is_same<Signed, NegativeValueT>::value
-                      ? 0x80000000ul
-                      : 0;  // Sign bits.
-
-  if (std::is_same<Signed, SignedValueT>::value) {
-    uint64_t a = bits & 0x8000000000000000ull;
-    sign = static_cast<uint32_t>(a >> 32);
-    bits = bits & 0x7FFFFFFFFFFFFFFFull;
-  }
-  if (IncludeZero) {
-    if (bits == 0u) return 0;
-  }
-
-  // Number of leading zeros is mapped to the exponent: 2^-clz
-  int clz = base_internal::CountLeadingZeros64(bits);
-  // Shift number left to erase leading zeros.
-  bits <<= IncludeZero ? clz : (clz & 63);
-  // Shift number right to remove bits that overflow double mantissa.  The
-  // direction of the shift depends on `clz`.
-  bits >>= (64 - FLT_MANT_DIG);
-
-  // Construct IEEE 754 float exponent.
-  // In the Signed case, bits is a 63-bit number with a 0 msb.  Adjust the
-  // exponent to account for that.
-  const uint32_t exp =
-      (std::is_same<Signed, SignedValueT>::value ? 127U : 126U) +
-      static_cast<uint32_t>(ExponentBias - clz);
-  constexpr int kExp = FLT_MANT_DIG - 1;
-  const uint32_t val = sign | (exp << kExp) | (bits & ((1U << kExp) - 1U));
-
-  float res;
-  static_assert(sizeof(res) == sizeof(val), "float is not 32 bit");
-  // Assumes that endian-ness is same for float and uint32_t.
-  std::memcpy(&res, &val, sizeof(res));
-
-  return res;
-}
-
-template <typename Result>
-struct RandU64ToReal {
-  template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-  static inline Result Value(uint64_t bits) {
-    return RandU64ToDouble<Signed, IncludeZero, ExponentBias>(bits);
-  }
-};
-
-template <>
-struct RandU64ToReal<float> {
-  template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-  static inline float Value(uint64_t bits) {
-    return RandU64ToFloat<Signed, IncludeZero, ExponentBias>(bits);
-  }
-};
-
-}  // namespace random_internal
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_
diff --git a/absl/random/internal/generate_real.h b/absl/random/internal/generate_real.h
new file mode 100644
index 00000000000..246d863e56e
--- /dev/null
+++ b/absl/random/internal/generate_real.h
@@ -0,0 +1,144 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
+#define ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
+
+// This file contains some implementation details which are used by one or more
+// of the absl random number distributions.
+
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <type_traits>
+
+#include "absl/base/internal/bits.h"
+#include "absl/meta/type_traits.h"
+#include "absl/random/internal/fastmath.h"
+#include "absl/random/internal/traits.h"
+
+namespace absl {
+namespace random_internal {
+
+// Tristate tag types controlling the output of GenerateRealFromBits.
+struct GeneratePositiveTag {};
+struct GenerateNegativeTag {};
+struct GenerateSignedTag {};
+
+// GenerateRealFromBits generates a single real value from a single 64-bit
+// `bits` with template fields controlling the output.
+//
+// The `SignedTag` parameter controls whether positive, negative,
+// or either signed/unsigned may be returned.
+//   When SignedTag == GeneratePositiveTag, range is U(0, 1)
+//   When SignedTag == GenerateNegativeTag, range is U(-1, 0)
+//   When SignedTag == GenerateSignedTag, range is U(-1, 1)
+//
+// When the `IncludeZero` parameter is true, the function may return 0 for some
+// inputs, otherwise it never returns 0.
+//
+// When a value in U(0,1) is required, use:
+//   Uniform64ToReal<double, PositiveValueT, true>;
+//
+// When a value in U(-1,1) is required, use:
+//   Uniform64ToReal<double, SignedValueT, false>;
+//
+//   This generates more distinct values than the mathematical equivalent
+//   `U(0, 1) * 2.0 - 1.0`.
+//
+// Scaling the result by powers of 2 (and avoiding a multiply) is also possible:
+//   GenerateRealFromBits<double>(..., -1);  => U(0, 0.5)
+//   GenerateRealFromBits<double>(..., 1);   => U(0, 2)
+//
+template <typename RealType,  // Real type, either float or double.
+          typename SignedTag = GeneratePositiveTag,  // Whether a positive,
+                                                     // negative, or signed
+                                                     // value is generated.
+          bool IncludeZero = true>
+inline RealType GenerateRealFromBits(uint64_t bits, int exp_bias = 0) {
+  using real_type = RealType;
+  using uint_type = absl::conditional_t<std::is_same<real_type, float>::value,
+                                        uint32_t, uint64_t>;
+
+  static_assert(
+      (std::is_same<double, real_type>::value ||
+       std::is_same<float, real_type>::value),
+      "GenerateRealFromBits must be parameterized by either float or double.");
+
+  static_assert(sizeof(uint_type) == sizeof(real_type),
+                "Mismatched unsinged and real types.");
+
+  static_assert((std::numeric_limits<real_type>::is_iec559 &&
+                 std::numeric_limits<real_type>::radix == 2),
+                "RealType representation is not IEEE 754 binary.");
+
+  static_assert((std::is_same<SignedTag, GeneratePositiveTag>::value ||
+                 std::is_same<SignedTag, GenerateNegativeTag>::value ||
+                 std::is_same<SignedTag, GenerateSignedTag>::value),
+                "");
+
+  static constexpr int kExp = std::numeric_limits<real_type>::digits - 1;
+  static constexpr uint_type kMask = (static_cast<uint_type>(1) << kExp) - 1u;
+  static constexpr int kUintBits = sizeof(uint_type) * 8;
+
+  int exp = exp_bias + int{std::numeric_limits<real_type>::max_exponent - 2};
+
+  // Determine the sign bit.
+  // Depending on the SignedTag, this may use the left-most bit
+  // or it may be a constant value.
+  uint_type sign = std::is_same<SignedTag, GenerateNegativeTag>::value
+                       ? (static_cast<uint_type>(1) << (kUintBits - 1))
+                       : 0;
+  if (std::is_same<SignedTag, GenerateSignedTag>::value) {
+    if (std::is_same<uint_type, uint64_t>::value) {
+      sign = bits & uint64_t{0x8000000000000000};
+    }
+    if (std::is_same<uint_type, uint32_t>::value) {
+      const uint64_t tmp = bits & uint64_t{0x8000000000000000};
+      sign = static_cast<uint32_t>(tmp >> 32);
+    }
+    // adjust the bits and the exponent to account for removing
+    // the leading bit.
+    bits = bits & uint64_t{0x7FFFFFFFFFFFFFFF};
+    exp++;
+  }
+  if (IncludeZero) {
+    if (bits == 0u) return 0;
+  }
+
+  // Number of leading zeros is mapped to the exponent: 2^-clz
+  // bits is 0..01xxxxxx. After shifting, we're left with 1xxx...0..0
+  int clz = base_internal::CountLeadingZeros64(bits);
+  bits <<= (IncludeZero ? clz : (clz & 63));  // remove 0-bits.
+  exp -= clz;                                 // set the exponent.
+  bits >>= (63 - kExp);
+
+  // Construct the 32-bit or 64-bit IEEE 754 floating-point value from
+  // the individual fields: sign, exp, mantissa(bits).
+  uint_type val =
+      (std::is_same<SignedTag, GeneratePositiveTag>::value ? 0u : sign) |
+      (static_cast<uint_type>(exp) << kExp) |
+      (static_cast<uint_type>(bits) & kMask);
+
+  // bit_cast to the output-type
+  real_type result;
+  memcpy(static_cast<void*>(&result), static_cast<const void*>(&val),
+         sizeof(result));
+  return result;
+}
+
+}  // namespace random_internal
+}  // namespace absl
+
+#endif  // ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
diff --git a/absl/random/internal/distribution_impl_test.cc b/absl/random/internal/generate_real_test.cc
similarity index 81%
rename from absl/random/internal/distribution_impl_test.cc
rename to absl/random/internal/generate_real_test.cc
index fcc169046d7..aa02f0c2c1b 100644
--- a/absl/random/internal/distribution_impl_test.cc
+++ b/absl/random/internal/generate_real_test.cc
@@ -12,57 +12,74 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "absl/random/internal/distribution_impl.h"
+#include "absl/random/internal/generate_real.h"
+
+#include <cfloat>
+#include <cstddef>
+#include <cstdint>
+#include <string>
 
 #include "gtest/gtest.h"
 #include "absl/base/internal/bits.h"
 #include "absl/flags/flag.h"
-#include "absl/numeric/int128.h"
 
 ABSL_FLAG(int64_t, absl_random_test_trials, 50000,
           "Number of trials for the probability tests.");
 
-using absl::random_internal::NegativeValueT;
-using absl::random_internal::PositiveValueT;
-using absl::random_internal::RandU64ToDouble;
-using absl::random_internal::RandU64ToFloat;
-using absl::random_internal::SignedValueT;
+using absl::random_internal::GenerateNegativeTag;
+using absl::random_internal::GeneratePositiveTag;
+using absl::random_internal::GenerateRealFromBits;
+using absl::random_internal::GenerateSignedTag;
 
 namespace {
 
-TEST(DistributionImplTest, U64ToFloat_Positive_NoZero_Test) {
+TEST(GenerateRealTest, U64ToFloat_Positive_NoZero_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<PositiveValueT, false>(a);
+    return GenerateRealFromBits<float, GeneratePositiveTag, false>(a);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), 2.710505431e-20f);
   EXPECT_EQ(ToFloat(0x0000000000000001), 5.421010862e-20f);
   EXPECT_EQ(ToFloat(0x8000000000000000), 0.5);
+  EXPECT_EQ(ToFloat(0x8000000000000001), 0.5);
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), 0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloat_Positive_Zero_Test) {
+TEST(GenerateRealTest, U64ToFloat_Positive_Zero_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<PositiveValueT, true>(a);
+    return GenerateRealFromBits<float, GeneratePositiveTag, true>(a);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), 0.0);
   EXPECT_EQ(ToFloat(0x0000000000000001), 5.421010862e-20f);
   EXPECT_EQ(ToFloat(0x8000000000000000), 0.5);
+  EXPECT_EQ(ToFloat(0x8000000000000001), 0.5);
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), 0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloat_Negative_NoZero_Test) {
+TEST(GenerateRealTest, U64ToFloat_Negative_NoZero_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<NegativeValueT, false>(a);
+    return GenerateRealFromBits<float, GenerateNegativeTag, false>(a);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), -2.710505431e-20f);
   EXPECT_EQ(ToFloat(0x0000000000000001), -5.421010862e-20f);
   EXPECT_EQ(ToFloat(0x8000000000000000), -0.5);
+  EXPECT_EQ(ToFloat(0x8000000000000001), -0.5);
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), -0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloat_Signed_NoZero_Test) {
+TEST(GenerateRealTest, U64ToFloat_Negative_Zero_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<SignedValueT, false>(a);
+    return GenerateRealFromBits<float, GenerateNegativeTag, true>(a);
+  };
+  EXPECT_EQ(ToFloat(0x0000000000000000), 0.0);
+  EXPECT_EQ(ToFloat(0x0000000000000001), -5.421010862e-20f);
+  EXPECT_EQ(ToFloat(0x8000000000000000), -0.5);
+  EXPECT_EQ(ToFloat(0x8000000000000001), -0.5);
+  EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), -0.9999999404f);
+}
+
+TEST(GenerateRealTest, U64ToFloat_Signed_NoZero_Test) {
+  auto ToFloat = [](uint64_t a) {
+    return GenerateRealFromBits<float, GenerateSignedTag, false>(a);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), 5.421010862e-20f);
   EXPECT_EQ(ToFloat(0x0000000000000001), 1.084202172e-19f);
@@ -72,9 +89,9 @@ TEST(DistributionImplTest, U64ToFloat_Signed_NoZero_Test) {
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), -0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloat_Signed_Zero_Test) {
+TEST(GenerateRealTest, U64ToFloat_Signed_Zero_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<SignedValueT, true>(a);
+    return GenerateRealFromBits<float, GenerateSignedTag, true>(a);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), 0);
   EXPECT_EQ(ToFloat(0x0000000000000001), 1.084202172e-19f);
@@ -84,9 +101,9 @@ TEST(DistributionImplTest, U64ToFloat_Signed_Zero_Test) {
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), -0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloat_Signed_Bias_Test) {
+TEST(GenerateRealTest, U64ToFloat_Signed_Bias_Test) {
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<SignedValueT, true, 1>(a);
+    return GenerateRealFromBits<float, GenerateSignedTag, true>(a, 1);
   };
   EXPECT_EQ(ToFloat(0x0000000000000000), 0);
   EXPECT_EQ(ToFloat(0x0000000000000001), 2 * 1.084202172e-19f);
@@ -96,9 +113,9 @@ TEST(DistributionImplTest, U64ToFloat_Signed_Bias_Test) {
   EXPECT_EQ(ToFloat(0xFFFFFFFFFFFFFFFF), 2 * -0.9999999404f);
 }
 
-TEST(DistributionImplTest, U64ToFloatTest) {
+TEST(GenerateRealTest, U64ToFloatTest) {
   auto ToFloat = [](uint64_t a) -> float {
-    return RandU64ToFloat<PositiveValueT, true>(a);
+    return GenerateRealFromBits<float, GeneratePositiveTag, true>(a);
   };
 
   EXPECT_EQ(ToFloat(0x0000000000000000), 0.0f);
@@ -150,44 +167,60 @@ TEST(DistributionImplTest, U64ToFloatTest) {
   }
 }
 
-TEST(DistributionImplTest, U64ToDouble_Positive_NoZero_Test) {
+TEST(GenerateRealTest, U64ToDouble_Positive_NoZero_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<PositiveValueT, false>(a);
+    return GenerateRealFromBits<double, GeneratePositiveTag, false>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), 2.710505431213761085e-20);
   EXPECT_EQ(ToDouble(0x0000000000000001), 5.42101086242752217004e-20);
   EXPECT_EQ(ToDouble(0x0000000000000002), 1.084202172485504434e-19);
   EXPECT_EQ(ToDouble(0x8000000000000000), 0.5);
+  EXPECT_EQ(ToDouble(0x8000000000000001), 0.5);
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), 0.999999999999999888978);
 }
 
-TEST(DistributionImplTest, U64ToDouble_Positive_Zero_Test) {
+TEST(GenerateRealTest, U64ToDouble_Positive_Zero_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<PositiveValueT, true>(a);
+    return GenerateRealFromBits<double, GeneratePositiveTag, true>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), 0.0);
   EXPECT_EQ(ToDouble(0x0000000000000001), 5.42101086242752217004e-20);
   EXPECT_EQ(ToDouble(0x8000000000000000), 0.5);
+  EXPECT_EQ(ToDouble(0x8000000000000001), 0.5);
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), 0.999999999999999888978);
 }
 
-TEST(DistributionImplTest, U64ToDouble_Negative_NoZero_Test) {
+TEST(GenerateRealTest, U64ToDouble_Negative_NoZero_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<NegativeValueT, false>(a);
+    return GenerateRealFromBits<double, GenerateNegativeTag, false>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), -2.710505431213761085e-20);
   EXPECT_EQ(ToDouble(0x0000000000000001), -5.42101086242752217004e-20);
   EXPECT_EQ(ToDouble(0x0000000000000002), -1.084202172485504434e-19);
   EXPECT_EQ(ToDouble(0x8000000000000000), -0.5);
+  EXPECT_EQ(ToDouble(0x8000000000000001), -0.5);
+  EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), -0.999999999999999888978);
+}
+
+TEST(GenerateRealTest, U64ToDouble_Negative_Zero_Test) {
+  auto ToDouble = [](uint64_t a) {
+    return GenerateRealFromBits<double, GenerateNegativeTag, true>(a);
+  };
+
+  EXPECT_EQ(ToDouble(0x0000000000000000), 0.0);
+  EXPECT_EQ(ToDouble(0x0000000000000001), -5.42101086242752217004e-20);
+  EXPECT_EQ(ToDouble(0x0000000000000002), -1.084202172485504434e-19);
+  EXPECT_EQ(ToDouble(0x8000000000000000), -0.5);
+  EXPECT_EQ(ToDouble(0x8000000000000001), -0.5);
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), -0.999999999999999888978);
 }
 
-TEST(DistributionImplTest, U64ToDouble_Signed_NoZero_Test) {
+TEST(GenerateRealTest, U64ToDouble_Signed_NoZero_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<SignedValueT, false>(a);
+    return GenerateRealFromBits<double, GenerateSignedTag, false>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), 5.42101086242752217004e-20);
@@ -198,9 +231,9 @@ TEST(DistributionImplTest, U64ToDouble_Signed_NoZero_Test) {
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), -0.999999999999999888978);
 }
 
-TEST(DistributionImplTest, U64ToDouble_Signed_Zero_Test) {
+TEST(GenerateRealTest, U64ToDouble_Signed_Zero_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<SignedValueT, true>(a);
+    return GenerateRealFromBits<double, GenerateSignedTag, true>(a);
   };
   EXPECT_EQ(ToDouble(0x0000000000000000), 0);
   EXPECT_EQ(ToDouble(0x0000000000000001), 1.084202172485504434e-19);
@@ -210,9 +243,9 @@ TEST(DistributionImplTest, U64ToDouble_Signed_Zero_Test) {
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), -0.999999999999999888978);
 }
 
-TEST(DistributionImplTest, U64ToDouble_Signed_Bias_Test) {
+TEST(GenerateRealTest, U64ToDouble_GenerateSignedTag_Bias_Test) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<SignedValueT, true, -1>(a);
+    return GenerateRealFromBits<double, GenerateSignedTag, true>(a, -1);
   };
   EXPECT_EQ(ToDouble(0x0000000000000000), 0);
   EXPECT_EQ(ToDouble(0x0000000000000001), 1.084202172485504434e-19 / 2);
@@ -222,9 +255,9 @@ TEST(DistributionImplTest, U64ToDouble_Signed_Bias_Test) {
   EXPECT_EQ(ToDouble(0xFFFFFFFFFFFFFFFF), -0.999999999999999888978 / 2);
 }
 
-TEST(DistributionImplTest, U64ToDoubleTest) {
+TEST(GenerateRealTest, U64ToDoubleTest) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<PositiveValueT, true>(a);
+    return GenerateRealFromBits<double, GeneratePositiveTag, true>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), 0.0);
@@ -296,9 +329,9 @@ TEST(DistributionImplTest, U64ToDoubleTest) {
   }
 }
 
-TEST(DistributionImplTest, U64ToDoubleSignedTest) {
+TEST(GenerateRealTest, U64ToDoubleSignedTest) {
   auto ToDouble = [](uint64_t a) {
-    return RandU64ToDouble<SignedValueT, false>(a);
+    return GenerateRealFromBits<double, GenerateSignedTag, false>(a);
   };
 
   EXPECT_EQ(ToDouble(0x0000000000000000), 5.42101086242752217004e-20);
@@ -379,10 +412,10 @@ TEST(DistributionImplTest, U64ToDoubleSignedTest) {
   }
 }
 
-TEST(DistributionImplTest, ExhaustiveFloat) {
+TEST(GenerateRealTest, ExhaustiveFloat) {
   using absl::base_internal::CountLeadingZeros64;
   auto ToFloat = [](uint64_t a) {
-    return RandU64ToFloat<PositiveValueT, true>(a);
+    return GenerateRealFromBits<float, GeneratePositiveTag, true>(a);
   };
 
   // Rely on RandU64ToFloat generating values from greatest to least when
diff --git a/absl/random/log_uniform_int_distribution.h b/absl/random/log_uniform_int_distribution.h
index ac43416ee57..956a69070cc 100644
--- a/absl/random/log_uniform_int_distribution.h
+++ b/absl/random/log_uniform_int_distribution.h
@@ -23,8 +23,8 @@
 #include <ostream>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
 #include "absl/random/internal/fastmath.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 #include "absl/random/internal/traits.h"
 #include "absl/random/uniform_int_distribution.h"
diff --git a/absl/random/poisson_distribution.h b/absl/random/poisson_distribution.h
index 7750b1c9ba5..23a953ff253 100644
--- a/absl/random/poisson_distribution.h
+++ b/absl/random/poisson_distribution.h
@@ -22,9 +22,9 @@
 #include <ostream>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
 #include "absl/random/internal/fast_uniform_bits.h"
 #include "absl/random/internal/fastmath.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 
 namespace absl {
@@ -164,9 +164,9 @@ typename poisson_distribution<IntType>::result_type
 poisson_distribution<IntType>::operator()(
     URBG& g,  // NOLINT(runtime/references)
     const param_type& p) {
-  using random_internal::PositiveValueT;
-  using random_internal::RandU64ToDouble;
-  using random_internal::SignedValueT;
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+  using random_internal::GenerateSignedTag;
 
   if (p.split_ != 0) {
     // Use Knuth's algorithm with range splitting to avoid floating-point
@@ -186,7 +186,8 @@ poisson_distribution<IntType>::operator()(
     for (int split = p.split_; split > 0; --split) {
       double r = 1.0;
       do {
-        r *= RandU64ToDouble<PositiveValueT, true>(fast_u64_(g));
+        r *= GenerateRealFromBits<double, GeneratePositiveTag, true>(
+            fast_u64_(g));  // U(-1, 0)
         ++n;
       } while (r > p.emu_);
       --n;
@@ -205,10 +206,11 @@ poisson_distribution<IntType>::operator()(
   // and k = max(f).
   const double a = p.mean_ + 0.5;
   for (;;) {
-    const double u =
-        RandU64ToDouble<PositiveValueT, false>(fast_u64_(g));  // (0, 1)
-    const double v =
-        RandU64ToDouble<SignedValueT, false>(fast_u64_(g));  // (-1, 1)
+    const double u = GenerateRealFromBits<double, GeneratePositiveTag, false>(
+        fast_u64_(g));  // U(0, 1)
+    const double v = GenerateRealFromBits<double, GenerateSignedTag, false>(
+        fast_u64_(g));  // U(-1, 1)
+
     const double x = std::floor(p.s_ * v / u + a);
     if (x < 0) continue;  // f(negative) = 0
     const double rhs = x * p.lmu_;
diff --git a/absl/random/uniform_int_distribution.h b/absl/random/uniform_int_distribution.h
index 02aa376555e..dc8ba8c1e2a 100644
--- a/absl/random/uniform_int_distribution.h
+++ b/absl/random/uniform_int_distribution.h
@@ -34,7 +34,6 @@
 #include <type_traits>
 
 #include "absl/base/optimization.h"
-#include "absl/random/internal/distribution_impl.h"
 #include "absl/random/internal/fast_uniform_bits.h"
 #include "absl/random/internal/iostream_state_saver.h"
 #include "absl/random/internal/traits.h"
diff --git a/absl/random/uniform_real_distribution.h b/absl/random/uniform_real_distribution.h
index 336abb3907b..bf2ed2c5ad3 100644
--- a/absl/random/uniform_real_distribution.h
+++ b/absl/random/uniform_real_distribution.h
@@ -39,8 +39,9 @@
 #include <limits>
 #include <type_traits>
 
-#include "absl/random/internal/distribution_impl.h"
+#include "absl/meta/type_traits.h"
 #include "absl/random/internal/fast_uniform_bits.h"
+#include "absl/random/internal/generate_real.h"
 #include "absl/random/internal/iostream_state_saver.h"
 
 namespace absl {
@@ -76,6 +77,7 @@ class uniform_real_distribution {
       // is not possible, so value generation cannot use the full range of the
       // real type.
       assert(range_ <= (std::numeric_limits<result_type>::max)());
+      assert(std::isfinite(range_));
     }
 
     result_type a() const { return lo_; }
@@ -151,10 +153,15 @@ template <typename URBG>
 typename uniform_real_distribution<RealType>::result_type
 uniform_real_distribution<RealType>::operator()(
     URBG& gen, const param_type& p) {  // NOLINT(runtime/references)
-  using random_internal::PositiveValueT;
+  using random_internal::GeneratePositiveTag;
+  using random_internal::GenerateRealFromBits;
+  using real_type =
+      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
+
   while (true) {
-    const result_type sample = random_internal::RandU64ToReal<
-        result_type>::template Value<PositiveValueT, true>(fast_u64_(gen));
+    const result_type sample =
+        GenerateRealFromBits<real_type, GeneratePositiveTag, true>(
+            fast_u64_(gen));
     const result_type res = p.a() + (sample * p.range_);
     if (res < p.b() || p.range_ <= 0 || !std::isfinite(p.range_)) {
       return res;