3eb94ec81b sync: Use decltype(auto) return type for WITH_LOCK (Carl Dong)
Pull request description:
> Now that we're using C++17, we can use the decltype(auto) return type
> for functions and lambda expressions.
>
> As demonstrated in this commit, this can simplify cases where previously
> the compiler failed to deduce the correct return type.
>
> Just for reference, for the "assign to ref" cases fixed here, there are
> 3 possible solutions:
>
> - Return a pointer and immediately deref as used before this commit
> - Make sure the function/lambda returns declspec(auto) as used after
> this commit
> - Class& i = WITH_LOCK(..., return std::ref(...));
>
> -----
>
> References:
> 1. https://en.cppreference.com/w/cpp/language/function#Return_type_deduction
> 2. https://en.cppreference.com/w/cpp/language/template_argument_deduction#Other_contexts
> 3. https://en.cppreference.com/w/cpp/language/auto
> 4. https://en.cppreference.com/w/cpp/language/decltype
>
> Explanations:
> 1. https://stackoverflow.com/a/21369192
> 2. https://stackoverflow.com/a/21369170
Thanks to sipa and ryanofsky for helping me understand this
ACKs for top commit:
jnewbery:
utACK 3eb94ec81b
hebasto:
ACK 3eb94ec81b, I have reviewed the code and it looks OK, I agree it can be merged. I have verified possible warnings:
ryanofsky:
Code review ACK 3eb94ec81b
Tree-SHA512: 5f55c7722aeca8ea70e5c1a8db93e93ba0e356e8967e7f607ada38003df4b153d73c29bd2cea8d7ec1344720d37d857ea7dbfd2a88da1d92e0e9cbb9abd287df
Previously, the validation_chainstatemanager_tests test suite
instantiated its own duplicate ChainstateManager on which tests were
performed.
This wasn't a problem for the specific actions performed in
that suite. However, the existence of this duplicate ChainstateManager
and the fact that many of our validation static functions reach for
g_chainman, ::Chain(state|)Active means we may end up acting on two
different CChainStates should we write more extensive tests in the
future.
This change adds a new ChainTestingSetup which performs all
initialization previously done by TestingSetup except:
1. RPC command registration
2. ChainState initialization
3. Genesis Activation
4. {Ban,Conn,Peer}Man initialization
Means that we will no longer need to initialize a duplicate
ChainstateManger in order to test the initialization codepaths of
CChainState and ChainstateManager.
Lastly, this change has the additional benefit of allowing for
review-only assertions meant to show correctness to work in future work
de-globalizing g_chainman.
In the test chainstatemanager_rebalance_caches, an additional
LoadGenesisBlock call is added as MaybeReblanaceCaches eventually calls
FlushBlockFile, which tries to access vinfoBlockFile[nLastBlockFile],
which is out of bounds when LoadGenesisBlock hasn't been called yet.
-----
Note for the future:
The class con/destructor inheritance structure we have for these
TestingSetup classes is probably not the most suitable abstraction. In
particular, for both TestingSetup and ChainTestingSetup, we need to stop
the scheduler first before anything else. Otherwise classes depending on
the scheduler may be referenced by the scheduler after said classes are
freed. This means that there's no clear parallel between our teardown
code and C++'s destructuring order for class hierarchies.
Future work should strive to coalesce (as much as possible) test and
non-test init codepaths and perhaps structure it in a more fail-proof
way.
Also adds CCoinsViewCache::ReallocateCache() to attempt to free
memory that the cacheCoins's allocator may be hanging onto when
downsizing the cache.
Adds `CChainState::m_coins{tip,db}_cache_size_bytes` data members
so that we can reference cache size on a per-chainstate basis for
flushing.
James O'Beirne