This avoids the following allocations over repeated use:
- A ``CharsetEncoder``
- A byte array to contain the contents of the string
- A wrapper, via ``ByteBuffer#wrap``, for the previously mentioned byte
array
This also removes a copy of the temporary byte array. This extra copy
is needed because ``String#getBytes`` returns a byte array and its
length must match the exact size of the contents.
This implementation requires that the builder retain a ``ByteBuffer``
and ``CharEncoder``. This is considered slower only for users that
will only allocate a single string in their buffers or none at all.
The improvement is especially potent if the builder is constantly reused
by its caller. This also eliminates the penalty for users that do not
use strings at all since the cost of allocating these fields is
now amortized.
The only allocation left in this code is a ``CharBuffer`` wrapper. It's
possible to eliminate this allocation using an additional copy but need
further profiling to see which is the bigger problem.
Removes the following allocations:
- ``CharsetDecoder`` is reused between calls
- ``CharBuffer#wrap`` removed in favor of heap
based char buffer that is reused
- Temporary ``char[]``, an intermediate copy inside ``StringCoding``
- Another ``char[]``, this is needed because ``StringCoding`` uses
a ``CharBuffer`` internally but returns a ``char[]``. Extra
characters need to be trimmed so this means yet another allocation
- Yet another ``char[]`` directly from ``__string`` for non-heap
based buffers
Removes the following copies
- No copy is performed to trim the allocation since a ``CharBuffer``
is used directly
- For non-heap based byte buffers, removes the copy that was
previously done in the __string function
This does need to get the TLS entry which implies at least some
contention on the thread object table and a fence.
A previous commit that added "f" for C++/Java/C# would break
on gcc of constants like 3f, which are now output as 3.0f
Tested: on Linux
Change-Id: If9cabbe3c6d6948a5050b8b123bda9c06e181f52
CreateUninitializedVector was performing the following actions:
1. call StartVector.
2. call make_space, and set buf to point to the reserved space.
3. call EndVector.
The problem is that a call to EndVector can ultimately call make_space, which
if the buffer is full, will cause a reallocation, invalidating the value stored
in buf. So setting buf needs to be delayed until after EndVector.
The following code, when run under valgrind shows a write to free'd memory before
the change, but no such error after:
int main()
{
flatbuffers::FlatBufferBuilder fbb(128);
char *buf = nullptr;
fbb.CreateUninitializedVector(128, &buf);
*buf = 0;
}
This overflow could allow malformed FlatBuffers to pass the verifier.
Change-Id: Ia993299a761b00b93e53e8aff2689b631755763c
Tested: on Linux.
Bug: 27903580
To support the use case described in issue google/flatbuffers#3826, a new command line option --gen-name-strings
has been added, which will cause a static GetFullyQualifiedName function to be added
to the C++ output for tables/structs.
Currently in order to get a value type of [ubyte] in PHP, iteration is
necessary which is not efficient. Helper __vector_as_bytes has been
implemented in order to return the byte arrays in PHP efficiently.
Appropriate methods also been added to use aforementioned method to return
the byte array. (e.g. get*Bytes()).
The reason the methods are named get*Bytes() and not for instance
get*ByteArray() is the fact that PHP doesn't support byte arrays and the
binary safe string implementation in PHP is used to simulate byte arrays
and since there is chance for PHP users to confuse this with PHP arrays
the name get*Bytes() has been chosen.
In the future __vector_as_bytebuffer() method can also be implemented to
return PHP implementation of ByteBuffer.
