Merge #16957: 9% less memory: make SaltedOutpointHasher noexcept

67d9990 make SaltedOutpointHasher noexcept (Martin Ankerl)

Pull request description:

  If the hash is not `noexcept`, `unorderd_map` has to assume that it can throw an exception. Thus when rehashing care needs to be taken. libstdc++ solves this by simply caching the hash value, which increases memory of each node by 8 bytes. Adding `noexcept` prevents this caching. In my experiments with `-reindex-chainstate -stopatheight=594000`, memory usage (maximum resident set size) has decreased by 9.4% while runtime has increased by 1.6% due to additional hashing. Additionally, memusage::DynamicUsage() is now more accurate and does not underestimate.

  |                                       | runtime h:mm:ss | max RSS kbyte |
  |---------------------------------------|-----------------|--------------|
  | master                                |         4:13:59 |      7696728 |
  | 2019-09-SaltedOutpointHasher-noexcept |         4:18:11 |      6971412 |
  | change                                |          +1.65% |       -9,42% |

  Comparison of progress masters vs. 2019-09-SaltedOutpointHasher-noexcept
  ![out](https://user-images.githubusercontent.com/14386/65541887-69424e00-df0e-11e9-8644-b3a068ed8c3f.png)

ACKs for top commit:
  jamesob:
    Tested ACK bitcoin/bitcoin@67d9990

Tree-SHA512: 9c44e3cca993b5a564dd61ebd2926b9c4a238609ea4d283514c018236f977d935e35a384dd4696486fd3d78781dd2ba190bb72596e20a5e931042fa465872a0b

Author: laanwj

src/coins.h

@@ -95,8 +95,16 @@ class SaltedOutpointHasher
      * This *must* return size_t. With Boost 1.46 on 32-bit systems the
      * unordered_map will behave unpredictably if the custom hasher returns a
      * uint64_t, resulting in failures when syncing the chain (#4634).
+     *
+     * Having the hash noexcept allows libstdc++'s unordered_map to recalculate
+     * the hash during rehash, so it does not have to cache the value. This
+     * reduces node's memory by sizeof(size_t). The required recalculation has
+     * a slight performance penalty (around 1.6%), but this is compensated by
+     * memory savings of about 9% which allow for a larger dbcache setting.
+     *
+     * @see https://gcc.gnu.org/onlinedocs/gcc-9.2.0/libstdc++/manual/manual/unordered_associative.html
      */
-    size_t operator()(const COutPoint& id) const {
+    size_t operator()(const COutPoint& id) const noexcept {
         return SipHashUint256Extra(k0, k1, id.hash, id.n);
     }
 };