FlexBuffers are used by custom op handlers for TFLite. Using the optimized path for ReadUInt64 in FlexBuffers causes a crash in models using custom ops on win32 build. This change fixes the problem by using unoptimized implementation of ReadUInt64 for win32.
ARM64EC is a new ARM64 ABI designed by Microsoft to support x64 application emulation on ARM64 CPUs. When compiling for ARM64EC, both the _M_X64 and _M_ARM64EC macros are defined. However, that causes problem in compiling this file, because the __movsb intrinsic, which is lowered to rep movsb, is not supported on ARM64, so the optimization for native x64 should be disabled for ARM64EC.
- add a new method ParseJson to minimize failures during fuzzing
- add default (conditional) move-constructor for Parser
- add a new monster_fuzzer
- switch fuzzers to C++17 and `test/cpp17` generated code
Add static cast from float to double in flexbuffers.h to avoid implicit double promotion error. This error is surfacing during the tensorflow lite for microcontrollers build since we enabled -Werror and -Wdouble-promotion.
Number of elements on the stack shouldn't affect the calculation
of ElemWidth(). Variable 'start' needs to be subtracted from the
loop variable 'i' to make indexing zero-based.
There is an additional unit test to pack nested vectors. Size of
the packed buffer *without* this fix is 798 and only 664 bytes
*with* the fix.
* Fixed refractoring issue in reflection/generate_code.sh. Also, mv deletes the original file, so I don't need to clean it up manually in that case.
* Added --filename-suffix and --filename-ext to flatc
* Fixed typo and added example generation of suffix and extension for C++
* Removed extra ;
* Removed clang-format block from a region that didn't need it. Fixed an auto format of another clang-format block
* Added docs, fixed pointer alignment, removed suffix test file
This means data written with older versions of this code has
potentially misaligned data, which we'll need to support.
This isn't a problem on most architectures, but could be on
older ARM chips. To support them properly may require swapping
out uses of `flatbuffers::ReadScalar` with a version that does a
memcpy internally.
Change-Id: Ib352aab4a586f3a8c6602fb25488dcfff61e06e0
* Automatic refractor of C++ headers to Google C++ style guide
* Automatic refractor of C++ source to Google C++ style guide
* Automatic refractor of C++ tests to Google C++ style guide
* Fixed clang-format issues by running clang-format twice to correct itself. Kotlin was missing clang-format on after turning it off, so it was changed,
With the introduction of Windows 10 on ARM (ARM64), code that assumes
that Windows targets are always x86 or x86_64 targets needs to be
updated.
The hot function ReadUInt64 has been optimized in MSVC builds using the
compiler intrinsic __movsb. Since this does not exist on ARM64 Windows,
this change uses the pure C++ path that other platforms use instead.
* FlexBuffer to JSON convertor for typed and fixedTypedvectors
* moving the common implementation to template
* signed unsigned comparison fix
* fix a formatting ({
* changing logic to append comma in vector of elements in json
* Add more apis to query vector types from a reference
https://github.com/google/flatbuffers/issues/4818
* changing order of apis
* another reordering
* removed vector element type api as not needed as for now
* Added support for the non-escaped print of utf-8 string.
* EscapeString: the first invalid symbol resets print_natural_utf8 flag to false.
* Move the test to ParseAndGenerateTextTest. Fixes.
* Removed dependence between `natural_utf8` and `allow_non_utf8` flags.
An assert in flexbuffers was bit-shifting a 64-bit number by
64 bits, which throws up warnings in some automated tools.
The same assert also checks to see if the number of bytes
being shifted is 8. Swapped the order, so that the bitshift
only occurs if the number of bits being shifted is not 64.
Should be the same behavior, but plays nicer with diagnostic
tools.
* Add support for boolean types in flexbuffers
* Simplify casting number <=> boolean
* Added comments for tests
* Add proper support for Booleans
* Bad rebase
* No special case for strings
* Removed unused test
* Simplify logic
* Added internal - limited - implementation of flatbuffers::unique_ptr
for STLs that don't ship with std::unique_ptr. In C++11 and beyond
this is just an alias for std::unique_ptr.
* Aliased used type traits structs is_scalar is_floating_point is_unsigned
into flatbuffers namespace so they can be replaced in C++98 implementations.
Right now these point at stlport's TR1 implementations.
* Wrapped vector::data() in vector_data().
* Wrapped vector::emplace_back() in vector_emplace_back().
* Wrapper string::back() in string_back().
* Added variants of FlatBufferBuilder::CreateVector() and
FlatBufferBuilder::CreateVectorOfStructs() that allow the use of plain
function pointers.
Generated code has also been modified to use plain functions to build objects
rather than std::function() so all generated code will work in C++98
applications.
* Added flexbuffers::Builder::Vector(), flexbuffers::Builder::TypedVector()
and flexbuffers::Builder::Map() methods that allow the use of plain function
pointers.
* Changed Parser to internally use plain function pointers when parsing table
and vector delimiters.
* Added specializations of NumToString() for 64-bit types that aren't supported
by stringstream in stlport.
* Overloaded numeric_limits for 64-bit types not supported by stlport.
* Replaced build_apk.sh (which was broken by deprecation of the
"android" tool in the Android SDK) with build.gradle and the
appropriate gradle wrapper to build an APK.
* Switched Android build to build against all STL variants.
* Updated travis configuration to build Android test and sample.
Tested:
* Verified all tests continue to work on Linux, OSX and Android.
* Verified Travis build is green.
Change-Id: I9e634363793f85b9f141d21454b10686020a2065
