* Vector of Tables equality
* support nullptr and fix for not being able to use auto in lambda
* use different std::equal overload
* use flatbuffers::unique_ptr
* go back to auto and clang-format fix
Augment the C++ generator to emit a C++ copy constructor and a by-value
copy assignment operator. This is enabled by default when the C++
standard is C++11 or later. These additional functions are only emitted
for objects that need it, typically tables containing other tables.
These new functions are declared in the object table type and are
defined as inline functions after table declarations.
When these new functions are declared, a user-defined
explicitly-defaulted default constructor and move constructor are also
emitted.
The copy assignment operator uses the copy-and-swap idiom to provide
strong exception safety, at the expense of keeping 2 full table copies
in memory temporarily.
fixes#5783
* CPP Default Value Generation in Mutators
If the mutator is for a value that is compatible with having a default value, then the single parameter becomes a default parameter. With this, a value can be mutated to it's default value without storing the default value, as that will be stored with the mutate function.
Fixed Casting When Generating Default for Enum Value
Added support for typecasting an int default value into the correct enum type in the default parameter. This fixed the issue of trying to use set a strongly typed enum parameter to an int which fails type checking.
Fixed Boolean Edge Case
Boolean types generate 0 != 0 when generating the underlying type which appears to be unique to the boolean type so it is now checked and the proper default value generated. It may be beneficial to check if it is instead an enum type, however the seeming edge case nature is why boolean was chosen to be checked.
Updated Generated Files
Regenerated the auto generated files to reflect the new changes.
Updated Remaining Files
Should fix auto generated header files that were not updated.
* Unified Repeated Code
Relocated identical append code to outside of conditional. Also changed 'casted' default value name from FIELD to INTERFACE to more accurately describe it.
* Moved Field Name Outside Conditional
Removed duplicate _{{FIELD_NAME}} and moved to unified append.
* flattests_cpp17 doesn't compile with Visual Studio 2017: warning C4100: 'indent': unreferenced formal parameter
stringify_util.h(127): error C2220: warning treated as error - no 'object' file generated
stringify_util.h(127): warning C4100: 'indent': unreferenced formal parameter
stringify_util.h(85): warning C4100: 'indent': unreferenced formal parameter
stringify_util.h(85): warning C4100: 'fbs': unreferenced formal parameter
* [C++] Add GetMutableSizePrefixedRoot() and generate a GetMutableSizePrefixed function
When using the mutable API together with size prefixed buffers these functions should be present.
* clang-format
* Cleanup branch for PR
Revert "flattests_cpp17 doesn't compile with Visual Studio 2017: warning C4100: 'indent': unreferenced formal parameter"
This reverts commit a92055203e.
Causes generated code to check for existing pointers in an object (when using the object based api) and use UnPackTo(existingPointer) rather than just using UnPack() to replace the pointer. This is a performance issue when unpacking to existing objects when using large buffer fields (e.g. image frames)
Co-authored-by: RobWauson <robwauson@gmail.com>
* Improve generated comparisons for tables
Previously, the generated code called std::unique_ptr<>::operator== on
non-scalar members, which would be correct only if the member was null
on both sides. After this CL, comparison is done on the pointed-to
values (including using a default-constructed value if the pointer is
null).
* Don't equate null Tables with all defaults
Also removes the cost of default-constructing tables for comparisons.
* Regenerate code
* fix formatting
Prior to this commit the default C++ code generator was `c++0x`.
A code generated with `c++0x` code-gen might have a vulnerability (undefined behavior) connected evolution of enums in a schema. This UB could break the backward compatibility if previously generated code casts an unknown enumerator to enum type that knows nothing about future enumerators added to the schema.
The main differences between `c++0x` and `c++11`:
- generated enums use explicitly declared underlying type;
- generated object-API tables don't declare default ctor() explicitly, instead of it default data member initializers are generated.
Please use `flatc --cpp-std c++0x` option for backward compatibility with old compilers.
UnpackTo copies vector elements one-by-one, which can be very inefficient depending on the quality of the compiler optimizations performed. This change updates the operation for vectors of bytes that aren't enums to use 'std::copy', which is usually highly optimized.
vectors of types that are more than one byte can't be optimized in this way because of the endianness of the serialized bytes vs. the target architecture endianness.
vectors of enums can't be optimized because they are required to be static_cast into the appropriate enum type when stored in the vector.
vectors of bools can be optimized in most cases, but since the standard
allows std::vector<bool> template specialization for space-savings,
std::copy doesn't work on every implementation (looking at you
Microsoft). Thus, this optimization is skipped for vector<bool>.
For a specific example, this improves the latency of unpacking large buffers on the Hexagon DSP by about 10x.
Co-authored-by: Matthew Markwell <markwell@google.com>
For example:
include/flatbuffers/reflection.h:365:8: error: definition of implicit copy
constructor for 'pointer_inside_vector<flatbuffers::Table, unsigned char>'
is deprecated because it has a user-declared copy assignment operator
[-Werror,-Wdeprecated-copy]
void operator=(const pointer_inside_vector &piv);
^
It's unclear why the old code wanted to declare a public `operator=`
without defining it; that just seems like a misunderstanding of the C++03 idiom
for deleting a member function. And anyway, we don't *want* to delete the
assignment operator; these are polymorphic types that do not follow value
semantics and nobody should ever be trying to copy them. So the simplest fix
is just to go back to the Rule of Zero: remove the declaration of `operator=`
and let the compiler do what it wanted to do originally anyway.
"The best code is no code."
Also, update the generated .h files.
Fixes#5649.
Previously UnPack would allocate data with new and assign it to a
raw pointer. This behavior makes it possible for the pointer to be
leaked in case of OOM. This commit defaults to use the user specified
pointer (which needs to implement a move constructor, a .get() and a
.release() operators), thus preventing these leaks.
* Add Builder and Table typedefs
This gives us a way to use templates to go from a builder to a table
and back again without having to pass both types in.
* Fix tests/cpp17/generated_cpp17/monster_test_generated.h
* Add flatc '--cpp_std' switch and sandbox for C++17 code generator
- Added 'flac --cpp_std legacy' for compatibility with old compilers (VS2010);
- Added experimental switch 'flac --cpp_std c++17' for future development;
- Added C++17 sandbox test_cpp17.cpp;
- C++ code generator generates enums with explicit underlying type to avoid problems with the forward and backward schema compatibility;
- Adjusted CMakeLists.txt, CI and generate code scripts to support of introduced '--cpp_std';
* Fix --cpp_std values: c++0x, c++11, c++17
* Add 'cpp::CppStandard' enum
* Add testing engine into test_cpp17
* Rebase to upstream/master
* Set default '--cpp-std C++0x'
* Fix code generation (--cpp_std C++11) in CMakeLists.txt
- Fix dependency declaration of grpctest target
* Revert --cpp-std for the tests from explicit C++11 to flatc default value (C++0x)
* Added a cpp UnPackSizePrefixed<struct_name> generated helper function
Missing helper function added to the cpp API generator for unpacking size prefixed structures
* Added generated test files
- update C++ monster_test::Color to unsigned type
- update Go Color:ubyte in the go_test.go
- add workaround for unsigned enum in java test
- sync generate.bat and generate.sh
* Fix the header file path in the tutorial doc
* Add the path field in sample/monster.fbs to match the tutorial
* Update the lobster sample file
* Update the binary sample file
Multiple calls of e.g. CreateString inside a call to a CreateTable
could cause those strings to end up in different locations in the
wire format, since order or argument evaluation is undefined.
This is allowed by the FlatBuffer format, but it is not helpful,
especially when debugging the contents of binaries, or comparing
against a "golden" binary for tests etc.
Now making sure that all the CreateTableDirect calls first serialize
sub strings/vectors before calling CreateTable.
Also made similar changes to the serialization of "binary schemas".
Change-Id: I5747c4038b37a0d400aca2bc592bec751cf5c172
Give the vtable offset enum inside each table the name
"FlatBuffersVTableOffset" and base type voffset_t so it can be used as a
dependent type in IsFieldPresent. This makes that function slightly
safer since it prevents calling it with arbitrary, non-table types.
Now, the only way to use IsFieldPresent incorrectly is to create your
own type which does not inherit from flatbuffers::Table but has a
dependent voffset convertible type "FlatBuffersVTableOffset".
Make an out-of-bounds check for enum values before using them to index the
names array. For consistency with non-sparse enums an empty string is
returned.
Fixes#4821
Derek Bailey