Add dragonbox to compute the required precision to print floating point
numbers. This avoids uglification of floating point numbers that
happen by default via std::stringstream.
Numbers like 34.34 used to be converted to '34.340000000000003' as strings.
With this version they will be converted to the string '34.34'.
This fixes issue https://github.com/jbeder/yaml-cpp/issues/1289
At the cost of exposing some existing detail, enable efficient
unsharing of graph structure (DeepClone), and efficient ShallowClone.
Modify[Key]Values allows in-place modification (or deletion) of
node_data in a Map. (the Key is not modified but provided so the
modification of the value may depend on the key)
UnshareSubtrees for in-place DeepClone copying of
shared subtrees.
id and hash_value allow creating efficient input Node -> output Node
transformations where the result of a previously transformed subtree
may be reused (allowing linear time creation of output trees with
shared subtrees instead of exponential by
Rationale: although YAML::Node faithfully represents the input
document and allows efficient traversal, using it as a large-scale
configuration / AST means that it's easy to unintentionally create
quadratic or worse algorithms modifying the tree.
The issue is that numbers like
2.01 or 3.01 can not be precisely represented with binary floating point
numbers.
This replaces all occurrences of 'std::numeric_limits<T>::digits10 + 1' with
'std::numeric_limits<T>::max_digits10'.
Background:
Using 'std::numeric_limits<T>::digits10 + 1' is not precise enough.
Converting a 'float' into a 'string' and back to a 'float' will not always
produce the original 'float' value. To guarantee that the 'string'
representation has sufficient precision the value
'std::numeric_limits<T>::max_digits10' has to be used.
Previously, just referencing the next element in the sequence (and so constructing it, as an undefined element) would allow you to skip defining an element without turning the sequence into a map. E.g:
node[0] = "foo"; // sequence of size 1
node[1]; // sequence of size 1, with an undefined element at 1
node[2] = "bar"; // FIX: should be map of size 2 (since there's no element at index 1)
* Change node_map type from map<ptr,ptr> to vector<pair<ptr,ptr>>
Map nodes are now iterated over in document order.
* Change insert_map_pair to always append
Always append in insert_map_pair even if the key is already present.
This breaks the behavior of force_insert which now always inserts KVs
even if the key is already present. The first insert for duplicated keys
now takes precedence for lookups.
"const Node Node::operator[](const Key& key) const" changed from
returning new empty node if the key was missing in 0.5.1 to returning
a shared 'zombie' node in 0.5.2 to resolve a memory leak.
(Specifically 1025f76df1 was where this
was introduced)
This caused some regressions where this 'zombie' object threw exceptions
in some functions where the 'empty' object would not.
This change fixes the Node::as(fallback) method (to return the
'fallback' instead of throwing an exception) and the
Node::begin()/Node::end() methods to return default-constructed
iterators (so begin() == end() in such cases) instead of another
exception.
- Update the call to equals() in node_data::remove() to match the new implementation
- Add unit test for node::remove() to catch this type of bug in the future
warning C4800: forcing value to bool 'true' or 'false' (performance warning)
for the node test, since it really doesn't make any sense in this context. (It's exactly what we intended with the "unspecified bool type".)
