Fix Randen and PCG on Big Endian platforms (#1031)

absl/random/internal/explicit_seed_seq.h

@@ -74,7 +74,7 @@ class ExplicitSeedSeq {
   template <typename OutIterator>
   void generate(OutIterator begin, OutIterator end) {
     for (size_t index = 0; begin != end; begin++) {
-      *begin = state_.empty() ? 0 : little_endian::FromHost32(state_[index++]);
+      *begin = state_.empty() ? 0 : state_[index++];
       if (index >= state_.size()) {
         index = 0;
       }

absl/random/internal/randen_engine.h

@@ -121,6 +121,13 @@ class alignas(16) randen_engine {
       const size_t requested_entropy = (entropy_size == 0) ? 8u : entropy_size;
       std::fill(std::begin(buffer) + requested_entropy, std::end(buffer), 0);
       seq.generate(std::begin(buffer), std::begin(buffer) + requested_entropy);
+#ifdef ABSL_IS_BIG_ENDIAN
+      // Randen expects the seed buffer to be in Little Endian; reverse it on
+      // Big Endian platforms.
+      for (sequence_result_type& e : buffer) {
+        e = absl::little_endian::FromHost(e);
+      }
+#endif
       // The Randen paper suggests preferentially initializing even-numbered
       // 128-bit vectors of the randen state (there are 16 such vectors).
       // The seed data is merged into the state offset by 128-bits, which

absl/random/internal/randen_slow.cc

@@ -395,6 +395,23 @@ inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void Permute(
   }
 }
 
+// Enables native loads in the round loop by pre-swapping.
+inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void SwapEndian(
+    absl::uint128* state) {
+#ifdef ABSL_IS_BIG_ENDIAN
+  for (uint32_t block = 0; block < RandenTraits::kFeistelBlocks; ++block) {
+    uint64_t new_lo = absl::little_endian::ToHost64(
+        static_cast<uint64_t>(state[block] >> 64));
+    uint64_t new_hi = absl::little_endian::ToHost64(
+        static_cast<uint64_t>((state[block] << 64) >> 64));
+    state[block] = (static_cast<absl::uint128>(new_hi) << 64) | new_lo;
+  }
+#else
+  // Avoid warning about unused variable.
+  (void)state;
+#endif
+}
+
 }  // namespace
 
 namespace absl {
@@ -439,8 +456,12 @@ void RandenSlow::Generate(const void* keys_void, void* state_void) {
 
   const absl::uint128 prev_inner = state[0];
 
+  SwapEndian(state);
+
   Permute(state, keys);
 
+  SwapEndian(state);
+
   // Ensure backtracking resistance.
   *state ^= prev_inner;
 }