Add FastRange32 function and use it throughout the codebase

Author: sipa

src/common/bloom.cpp

@@ -11,6 +11,7 @@
 #include <script/standard.h>
 #include <span.h>
 #include <streams.h>
+#include <util/fastrange.h>
 
 #include <algorithm>
 #include <cmath>
@@ -191,14 +192,6 @@ static inline uint32_t RollingBloomHash(unsigned int nHashNum, uint32_t nTweak,
     return MurmurHash3(nHashNum * 0xFBA4C795 + nTweak, vDataToHash);
 }
 
-
-// A replacement for x % n. This assumes that x and n are 32bit integers, and x is a uniformly random distributed 32bit value
-// which should be the case for a good hash.
-// See https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
-static inline uint32_t FastMod(uint32_t x, size_t n) {
-    return ((uint64_t)x * (uint64_t)n) >> 32;
-}
-
 void CRollingBloomFilter::insert(Span<const unsigned char> vKey)
 {
     if (nEntriesThisGeneration == nEntriesPerGeneration) {
@@ -223,7 +216,7 @@ void CRollingBloomFilter::insert(Span<const unsigned char> vKey)
         uint32_t h = RollingBloomHash(n, nTweak, vKey);
         int bit = h & 0x3F;
         /* FastMod works with the upper bits of h, so it is safe to ignore that the lower bits of h are already used for bit. */
-        uint32_t pos = FastMod(h, data.size());
+        uint32_t pos = FastRange32(h, data.size());
         /* The lowest bit of pos is ignored, and set to zero for the first bit, and to one for the second. */
         data[pos & ~1] = (data[pos & ~1] & ~(((uint64_t)1) << bit)) | ((uint64_t)(nGeneration & 1)) << bit;
         data[pos | 1] = (data[pos | 1] & ~(((uint64_t)1) << bit)) | ((uint64_t)(nGeneration >> 1)) << bit;
@@ -235,7 +228,7 @@ bool CRollingBloomFilter::contains(Span<const unsigned char> vKey) const
     for (int n = 0; n < nHashFuncs; n++) {
         uint32_t h = RollingBloomHash(n, nTweak, vKey);
         int bit = h & 0x3F;
-        uint32_t pos = FastMod(h, data.size());
+        uint32_t pos = FastRange32(h, data.size());
         /* If the relevant bit is not set in either data[pos & ~1] or data[pos | 1], the filter does not contain vKey */
         if (!(((data[pos & ~1] | data[pos | 1]) >> bit) & 1)) {
             return false;

src/cuckoocache.h

@@ -5,6 +5,8 @@
 #ifndef BITCOIN_CUCKOOCACHE_H
 #define BITCOIN_CUCKOOCACHE_H
 
+#include <util/fastrange.h>
+
 #include <algorithm> // std::find
 #include <array>
 #include <atomic>
@@ -219,13 +221,8 @@ class cache
      * One option would be to implement the same trick the compiler uses and compute the
      *  constants for exact division based on the size, as described in "{N}-bit Unsigned
      *  Division via {N}-bit Multiply-Add" by Arch D. Robison in 2005. But that code is
-     *  somewhat complicated and the result is still slower than other options:
-     *
-     * Instead we treat the 32-bit random number as a Q32 fixed-point number in the range
-     *  [0, 1) and simply multiply it by the size. Then we just shift the result down by
-     *  32-bits to get our bucket number. The result has non-uniformity the same as a
-     *  mod, but it is much faster to compute. More about this technique can be found at
-     *  https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/ .
+     *  somewhat complicated and the result is still slower than an even simpler option:
+     *  see the FastRange32 function in util/fastrange.h.
      *
      * The resulting non-uniformity is also more equally distributed which would be
      *  advantageous for something like linear probing, though it shouldn't matter
@@ -241,14 +238,14 @@ class cache
      */
     inline std::array<uint32_t, 8> compute_hashes(const Element& e) const
     {
-        return {{(uint32_t)(((uint64_t)hash_function.template operator()<0>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<1>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<2>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<3>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<4>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<5>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<6>(e) * (uint64_t)size) >> 32),
-                 (uint32_t)(((uint64_t)hash_function.template operator()<7>(e) * (uint64_t)size) >> 32)}};
+        return {{FastRange32(hash_function.template operator()<0>(e), size),
+                 FastRange32(hash_function.template operator()<1>(e), size),
+                 FastRange32(hash_function.template operator()<2>(e), size),
+                 FastRange32(hash_function.template operator()<3>(e), size),
+                 FastRange32(hash_function.template operator()<4>(e), size),
+                 FastRange32(hash_function.template operator()<5>(e), size),
+                 FastRange32(hash_function.template operator()<6>(e), size),
+                 FastRange32(hash_function.template operator()<7>(e), size)}};
     }
 
     /** invalid returns a special index that can never be inserted to

src/util/fastrange.h

@@ -7,11 +7,21 @@
 
 #include <cstdint>
 
-// Map a value x that is uniformly distributed in the range [0, 2^64) to a
-// value uniformly distributed in [0, n) by returning the upper 64 bits of
-// x * n.
-//
-// See: https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
+/* This file offers implementations of the fast range reduction technique described
+ * in https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
+ *
+ * In short, they take an integer x and a range n, and return the upper bits of
+ * (x * n). If x is uniformly distributed over its domain, the result is as close to
+ * uniformly distributed over [0, n) as (x mod n) would be, but significantly faster.
+ */
+
+/** Fast range reduction with 32-bit input and 32-bit range. */
+static inline uint32_t FastRange32(uint32_t x, uint32_t n)
+{
+    return (uint64_t{x} * n) >> 32;
+}
+
+/** Fast range reduction with 64-bit input and 64-bit range. */
 static inline uint64_t FastRange64(uint64_t x, uint64_t n)
 {
 #ifdef __SIZEOF_INT128__