Here is wisdom about how to build, test and run simdjson from within the repository. This is mostly useful for people who plan to contribute simdjson, or maybe study the design.
- Stage 1. (Find marks) Identifies quickly structure elements, strings, and so forth. We validate UTF-8 encoding at that stage.
- Stage 2. (Structure building) Involves constructing a "tree" of sort (materialized as a tape) to navigate through the data. Strings and numbers are parsed at this stage.
The role of stage 1 is to identify pseudo-structural characters as quickly as possible. A character is pseudo-structural if and only if:
1. Not enclosed in quotes, AND
2. Is a non-whitespace character, AND
3. Its preceding character is either:
(a) a structural character, OR
(b) whitespace OR
(c) the final quote in a string.
This helps as we redefine some new characters as pseudo-structural such as the characters 1, G, n in the following:
> { "foo" : 1.5, "bar" : 1.5 GEOFF_IS_A_DUMMY bla bla , "baz", null }
Stage 1 also does unicode validation.
Stage 2 handles all of the rest: number parsings, recognizing atoms like true, false, null, and so forth.
* **.github/workflows:** Definitions for GitHub Actions (CI).
* **singleheader:** Contains generated `simdjson.h` and `simdjson.cpp` that we release. The files `singleheader/simdjson.h` and `singleheader/simdjson.cpp` should never be edited by hand.
for it. Many of our benchmarks are microbenchmarks. We are effectively doing controlled scientific experiments for the purpose of understanding what affects our performance. So we simplify as much as possible. We try to avoid irrelevant factors such as page faults, interrupts, unnnecessary system calls. We recommend checking the performance as follows:
The last line becomes `./benchmark/Release/bench_parse_call.exe` under Windows. Under Windows, you can also build with the clang compiler by adding `-T ClangCL` to the call to `cmake ..`: `cmake .. - TClangCL`.
* **tools:** Source for executables that can be distributed with simdjson. Some examples:
*`json2json mydoc.json` parses the document, constructs a model and then dumps back the result to standard output.
*`json2json -d mydoc.json` parses the document, constructs a model and then dumps model (as a tape) to standard output. The tape format is described in the accompanying file `tape.md`.
*`minify mydoc.json` minifies the JSON document, outputting the result to standard output. Minifying means to remove the unneeded white space characters.
*`jsonpointer mydoc.json <jsonpath><jsonpath> ... <jsonpath>` parses the document, constructs a model and then processes a series of [JSON Pointer paths](https://tools.ietf.org/html/rfc6901). The result is itself a JSON document.
A key feature of simdjson is the ability to compile different processing kernels, optimized for specific instruction sets, and to select
the most appropriate kernel at runtime. This ensures that users get the very best performance while still enabling simdjson to run everywhere.
This technique is frequently called runtime dispatching. The simdjson achieves runtime dispatching entirely in C++: we do not assume
that the user is building the code using CMake, for example.
To make runtime dispatching work, it is critical that the code be compiled for the lowest supported processor. In particular, you should
not use flags such as -mavx2, /arch:AVX2 and so forth while compiling simdjson. When you do so, you allow the compiler to use advanced
instructions. In turn, these advanced instructions present in the code may cause a runtime failure if the runtime processor does not
support them. Even a simple loop, compiled with these flags, might generate binary code that only run on advanced processors.
So we compile simdjson for a generic processor. Our users should do the same if they want simdjson's runtime dispatch to work. It is important
to understand that if runtime dispatching does not work, then simdjson will cause crashes on older processors. Of course, if a user chooses
to compile their code for a specific instruction set (e.g., AVX2), they are responsible for the failures if they later run their code
on a processor that does not support AVX2. Yet, if we were to entice these users to do so, we would share the blame: thus we carefully instruct
users to compile their code in a generic way without doing anything to enable advanced instructions.
We only use runtime dispatching on x64 (AMD/Intel) platforms, at the moment. On ARM processors, we would need a standard way to query, at runtime,
the processor for its supported features. We do not know how to do so on ARM systems in general. Thankfully it is not yet a concern: 64-bit ARM
processors are fairly uniform as far as the instruction sets they support.
In all cases, simdjson uses advanced instructions by relying on "intrinsic functions": we do not write assembly code. The intrinsic functions
are special functions that the compiler might recognize and translate into fast code. To make runtime dispatching work, we rely on the fact that
the header providing these instructions
(intrin.h under Visual Studio, x86intrin.h elsewhere) defines all of the intrinsic functions, including those that are not supported
processor.
At this point, we are require to use one of two main strategies.
1. On POSIX systems, the main compilers (LLVM clang, GNU gcc) allow us to use any intrinsic function after including the header, but they fail to inline the resulting instruction if the target processor does not support them. Because we compile for a generic processor, we would not be able to use most intrinsic functions. Thankfully, more recent versions of these compilers allow us to flag a region of code with a specific target, so that we can compile only some of the code with support for advanced instructions. Thus in our C++, one might notice macros like `TARGET_HASWELL`. It is then our responsability, at runtime, to only run the regions of code (that we call kernels) matching the properties of the runtime processor. The benefit of this approach is that the compiler not only let us use intrinsic functions, but it can also optimize the rest of the code in the kernel with advanced instructions we enabled.
2. Under Visual Studio, the problem is somewhat simpler. Visual Studio will not only provide the intrinsic functions, but it will also allow us to use them. They will compile just fine. It is at runtime that they may cause a crash. So we do not need to mark regions of code for compilation toward advanced processors (e.g., with `TARGET_HASWELL` macros). The downside of the Visual Studio approach is that the compiler is not allowed to use advanced instructions others than those we specify. In principle, this means that Visual Studio has weaker optimization opportunities.
We also handle the special case where a user is compiling using LLVM clang under Windows, [using the Visual Studio toolchain](https://devblogs.microsoft.com/cppblog/clang-llvm-support-in-visual-studio/). If you compile with LLVM clang under Visual Studio, then the header files (intrin.h or x86intrin.h) no longer provides the intrinsic functions that are unsupported by the processor. This appears to be deliberate on the part of the LLVM engineers. With a few lines of code, we handle this scenario just like LLVM clang under a POSIX system, but forcing the inclusion of the specific headers, and rolling our own intrinsic function as needed.
In some cases, you may want to specify your compiler, especially if the default compiler on your system is too old. You need to tell cmake which compiler you wish to use by setting the CC and CXX variables. Under bash, you can do so with commands such as `export CC=gcc-7` and `export CXX=g++-7`. You can also do it as part of the `cmake` command: `cmake .. -DCMAKE_CXX_COMPILER=g++`. You may proceed as follows:
If your compiler does not default on C++11 support or better you may get failing tests. If so, you may be able to exclude the failing tests by replacing `ctest` with `ctest -E "^quickstart$"`.
Note that the name of directory (`build`) is arbitrary, you can name it as you want (e.g., `buildgcc`) and you can have as many different such directories as you would like (one per configuration).
- Grab the simdjson code from GitHub, e.g., by cloning it using [GitHub Desktop](https://desktop.github.com/).
- Install [CMake](https://cmake.org/download/). When you install it, make sure to ask that `cmake` be made available from the command line. Please choose a recent version of cmake.
- This last command (`cmake ...`) created a Visual Studio solution file in the newly created directory (e.g., `simdjson.sln`). Open this file in Visual Studio. You should now be able to build the project and run the tests. For example, in the `Solution Explorer` window (available from the `View` menu), right-click `ALL_BUILD` and select `Build`. To test the code, still in the `Solution Explorer` window, select `RUN_TESTS` and select `Build`.
Though having Visual Studio installed is necessary, one can build simdjson using only cmake commands:
-`mkdir build`
-`cd build`
-`cmake ..`
-`cmake --build . -config Release`
Furthermore, if you have installed LLVM clang on Windows, for example as a component of Visual Studio 2019, you can configure and build simdjson using LLVM clang on Windows using cmake:
