Compile-time JSON deserialization in C++(medium.com) |
Compile-time JSON deserialization in C++(medium.com) |
You do have to tag struct fields with a macro, but you can attach contexpr-visitable attributes. There's also a static limit to how many reflectable fields you can have, all reflectable fields need to be at the front of the struct, and the struct needs to be an aggregate.
Now I use the same trick in our code base to generically hash aggregates, but I limit it to 4 fields for sanity.
For this kind of things I tend to prefer using a simpler program (written in anything you like) to generate C or C++ instead of having the compile do the same thing much slowly.
Meta programming can be good, but it is even better done with an actual meta program, IMO.
Even though compilation time is the bane of C++, I think this concern regarding this specific usage is grossly overblown. I'm going to tell you why.
With incremental builds you only rebuild whatever has changed in your project. Embedding JSON documents in a C++ app is the kind of thing that is rarely touched, specially if all your requirements are met by serializing docs at compile time. This means that this deserialization will only be rarely rebuilt, and only under two scenarios: full rebuild, and touching the file.
As far as full rebuilds go, there is no scenario where deserializing JSON represents a relevant task in your build tree.
As for touching the file, if for some weird and unbelievable reason the build step for the JSON deserialization component is deemed too computationally expensive, it's trivial to move this specific component into a subproject that's built independently. This means that the full cost of an incremental build boils down to a) rebuilding your tiny JSON deserialization subproject, b) linking. Step a) runs happily in parallel with any other build task, thus it's impact is meaningless.
To read more on the topic, google for "horizontal architecture", a concept popularized by the book "Large-Scale C++: Process and Architecture, Volume 1" By John Lakos.
Mountain out of a molehill.
Like, for this particular example, you might start out with a header that looks like:
SomeData get_data_from_json(std::string_view json);
Then somebody comes around and says "we'd like to reuse the parsing logic to get SomeOtherData as well" and your nice, one-line header becomes
template<typename Ret>
Ret get_data_from_json(std::string_view json) {
// .. a gazillion lines of template-heavy code
}
It takes more discipline than you think across a team to prevent this from happening, mostly because a lot of people don't take "this takes too long to compile" as a serious complaint if it involves any kind of other trade-off.
That's a wild-assed guess. A JSON decoder right in the compiler could easily be faster than generation involving extra tool invocations and multiple passes.
Also, if you use ten code generators for ten different features in a pipeline instead of ten compile-time things built into the language, will that still be faster? What if most files use just use one one or two features? You have to pass them through all the generators just in case; each generator decides whether the file contains anything that it knows how to expand.
The C# approach for this is that code generators operate as compiler plugins (and therefore also IDE plugins, so if you report an error from the code generator it goes with all the other compile errors). There is a two-pass approach where your plugin gets to scan the syntax tree quickly for "might be relevant" and then another go later; the first pass is cached.
A limitation of the plugin approach is that your codegen code itself has to be in a separate project that gets compiled first.
An argument in favor of separate-codegen is that if it breaks you can inspect the intermediate code, and indeed do things like breakpoints, logging and inspection in the code generator itself. The C++ approach seems like it might be hard to debug in some situations.
It also can easily be slower: C++ templates are not exactly known for their blazingly fast compilation speed. Besides, the program they encode in this case is effectively being interpreted by the C++ compiler which, I suppose, is not really optimized for that: it's still mostly oriented around emitting optimized machine code.
Me too. The best example I can come up with is loading test data in automated tests, but even then I wouldn't use this sort of approach.
Another advantage of using KSP is that it also handles caching for you and will avoid running code generators again if the output already exists.
Still better than Rust /s
The advantage being that the parser is tailored to the specific type that is deserialized and it writes directly to the struct's fields instead of going through some dictionary.
template <StringLiteral str> static constexpr Key<str> key
in the class namespace - I think this this would work, but if I understand this correctly, You would need to do like
User user {...}; user[User::key<"myKey">]
Which actually is not so bad...
But please don’t do this in production.
What ever you need to do, use C++ templates as the last resort because you’ve figured out all other approaches suck even more. Maintaining template heavy code is absolutely horrible and wasteful (and if it’s C++ production code we measure it’s lifetime in decades). And no, there is no way ”to do it correctly so it doesn’t suck”.
Templates belong to the lowest abstraction levels - as stl mostly does. Anyhting more prevalent is an abomination.
If the schema is fixed, have types with the data and if you have a default data, provide it using initializer lists.
Ie. have a struct or structs with explicit serializeToJson and deserializeFromJson functions.
It’s faster to write than figuring out the correct template gymnastics and about 100x easier to maintain and extend.
Wrap everything in types and all is fine.
Verbose C++ is the only good kind of C++. Why?
C++ code needs to be debuggable and modifiable so when a profiler shows hotspots, you know where they are coming from and react appropriately.
How do you fix a hotspot on a one line of code somewhere in the middle of a template thingsmajic?
You don’t. You need to unroll the template code to untemplated code and the fix the hotspot.
Cases where you don’t need to fix one line hotspots because the resources consumed by the code are irrelevant are fine. But if performance does not matter it likely means you should use a bette language than C++.
Using C++ and not caring about performance finetuning is the worst of both worlds - you are using a cumbersome language AND it’s not even for any practical benefit.
(defmacro macro-time (&rest forms)
`(quote ,(eval `(progn ,@forms))))
For instance, if we have a snarf-file function which reads a text file and returns the contents as a string, we can do:
(macro-time (snarf-file "foo.txt"))
Seeing Greenspun's tenth rule [1] in action again and again is one of the weird things we Common Lisp programmers have to endure. I wish we would have more discussions on how to improve Lisp even further instead of trying to 'fix' C or C++ for the umpteenth time.
I agree one million percent; projects like SBCL are great, but my impression is that there are tons of improvements to be had in producing optimized code for modern processors (cache friendliness, SIMD, etc), GPU programming etc. I asked about efforts in those directions here and there, but did not get very clear answers.
If you're looking for an interesting follow-up project, here's something I had to do once that's now become much easier: compute a compile-time hash of the compilation for the current translation unit, e.g. __BASE_FILE__ hashed together with __TIMESTAMP__ or the equivalents for each platform.
This allows you to dynamically invalidate on-disk caches and trigger new-build tripwires based on ongoing revisions. Development and release builds are handled identically: if source file X handles a cache and X was recompiled, discard the cache.
I've written 2 iterations of a reflection library where you needed to annotate structs slightly with an ugly macro but once done you could just do: Message msg; if (parse_json(str,msg)) { ..process msg struct.. }
The previous iterations were for C++11 and C++17 but it seems that with C++20 features you don't even seem to need the macro uglyness so I personally think libraries need to move in the direction of plain old structs.
I did not see any mention of this in the post; so are you actually simply extracting the string versions of the numbers, without verifying nor deserializing them?
long double result = 0.0;
while (...) {
if (json[head] == '.') ...
result *= 10; result += json[head] - '0';
in a constexpr function with no problem :)
In terms of the implementation ... I feel like C++ is best when used in an "orthodox style" and minimizing the use of templates as much as possible.
--
1: https://learn.microsoft.com/en-us/dotnet/fsharp/tutorials/ty...
In JSON Link's case, since it was using C++17 at the time, it forced me to think around the problem of allocation and who does it. The library does not allocate, but potentially the data structures being deserialized to will. In C++20 you can get limited constexpr allocations but they are good for things like stacks and eliminating the fixed buffers many devs have used in the past; which is a good thing on it's own but isn't really allowing one to parse to a vector at compile time(as in OP's example) for things that persist.
Where this will get really interesting, though, is when #embed is in the major compilers. It's mostly there in clang, with gcc on the way I believe. It will open the door for DSL's and compile time configs in human readable formats or interop with other tools(maybe GUI designers)
As for OP's library, I am not a fan of the json_value like library approach that treats JSON as a thing to care about when it is usually just an imp detail to move to ones business objects.
TL;DR The big benefit though, is the ability to reason about the quality of the code in the library and have stronger testing.
I'm curious since I've gone back and forth on this in my own career. Both approaches come with their own pros and cons, but each get us to the same place.
Thanks for reading the article though, that's cool. I am a daw_json_link fan
Obviously, you have files that are not generated. They don't need any gen tool.
That's a disadvantage. If you want to start using JSON at compile time in a file, and the technology for that is a code generator, you have to move that file to a different category, perhaps by changing its suffix, and possibly indicate it somewhere in the build system as one of the sources needing the json generator. Whereas if it's in the language, you just do it in your .cpp file and that's it.
Token based macro preprocessors and code generators are simply not defensible in the face of structural macro systems and compile-time evaluation. They are just something you use when you don't have the latter. You can use code generators and preprocessors with languages that don't have anything built in, and which are resistant to change (will not support any decent metaprogramming in the foreseeable future).
I don't think you could parse it with, say, a class that has a std::string member (because of the transience restriction), but perhaps you can use lambdas that capture that string by reference, and call each other as appropriate?
As for exporting that as some sort of compiler artefact for use elsewhere, I am not sure how you would do that...
One place I have used the compile time, since JSON Link uses declarative mappings and is type based, is allowing for automatic (de)serialization of function calls over RPC/IPC. This makes things like communicating with web engines really neat, or doing serialized RPC calls over some other mechanism. Code like `server.register( "path", [] ( int x, Foo foo ) { ... } );` is working now in C++, and that's really neat and closely analogues some of the higher level API's in other languages.
Compared to what alternative that does the same thing, in C++?
Modern C++ compilers, as such, are slow as molasses, on multi-GHz hardware with huge L1 and L2 caches, whether your code uses templates or not.
I googled and found two examples of constexpr float parsing repositories, but from the sounds of things, you understand this problem better than I and will have seen them already
* (defun fx-add (x y)
(declare (optimize (speed 3) (safety 0) (debug 0))
(type fixnum x y))
(+ x y))
FX-ADD
* (disassemble #'fx-add)
; disassembly for FX-ADD
; Size: 104 bytes. Origin: #x7005970068 ; FX-ADD
; 68: 40FD4193 ASR NL0, R0, #1
; 6C: 00048B8B ADD NL0, NL0, R1, ASR #1
; 70: 0A0000AB ADDS R0, NL0, NL0
; 74: E7010054 BVC L1
; 78: BD2A00B9 STR WNULL, [THREAD, #40] ; pseudo-atomic-bits
; 7C: A97A47A9 LDP TMP, LR, [THREAD, #112] ; mixed-tlab.{free-pointer, end-addr}
; 80: 2A410091 ADD R0, TMP, #16
; 84: 5F011EEB CMP R0, LR
; 88: C8010054 BHI L2
; 8C: AA3A00F9 STR R0, [THREAD, #112] ; mixed-tlab
; 90: L0: 2A3D0091 ADD R0, TMP, #15
; 94: 3E2280D2 MOVZ LR, #273
; 98: 3E0100A9 STP LR, NL0, [TMP]
; 9C: BF3A03D5 DMB ISHST
; A0: BF2A00B9 STR WZR, [THREAD, #40] ; pseudo-atomic-bits
; A4: BE2E40B9 LDR WLR, [THREAD, #44] ; pseudo-atomic-bits
; A8: 5E0000B4 CBZ LR, L1
; AC: 200120D4 BRK #9 ; Pending interrupt trap
; B0: L1: FB031AAA MOV CSP, CFP
; B4: 5A7B40A9 LDP CFP, LR, [CFP]
; B8: BF0300F1 CMP NULL, #0
; BC: C0035FD6 RET
; C0: L2: 090280D2 MOVZ TMP, #16
; C4: 2AFCFF58 LDR R0, #x7005970048 ; SB-VM::ALLOC-TRAMP
; C8: 40013FD6 BLR R0
; CC: F1FFFF17 B L0
NIL
Furthermore, I’ve also decided to evaluate the debuggers available for Common Lisp. However, despite it being touted as a debugger-oriented language, I think the actual debuggers are pretty subpar, compared to debuggers available for C, C++, Java or .NET. No Common Lisp debugger supports watchpoints of any kind. If a given debugger supports breakpoints at all, they’re often done through wrapping code in code that triggers a breakpoint, or making this code run under interpreter instead of being native. Setting breakpoints in arbitrary code won’t work, it needs to be available as source code first. SBCL with SLIME doesn’t have a nice GUI where I could use the standard F[N] keys to step, continue, stop, etc. I don’t see any pane with live disassembly view. No live watch. LispWorks GUI on the other hand looks like a space station, where I struggle to orient myself. The only feature that is somewhat well-done is live code reload, but IMO it’s something far less important than well-implemented breakpoints and watchpoints in other languages, since the main thing I need the debugger for is to figure out what the hell a given piece of code is doing. Editing it is a completely secondary concern. And live code reload is also not unique to Common Lisp.
Debugger-wise, Java and .NET seem to be leading in quality, followed by C and C++.
[1]: Yes, I have read many comments about the alleged good performance of Common Lisp, but either authors of these comments live in a parallel reality with completely different benchmark results, or they’re comparing to Python. As such I treat those comments as urban legends.
* (defun fx-add (x y)
(declare (optimize (speed 3) (safety 0) (debug 0))
(type fixnum x y))
(the fixnum (+ x y)))
FX-ADD
* (disassemble 'fx-add)
; disassembly for FX-ADD
; Size: 6 bytes. Origin: #x552C81A6 ; FX-ADD
; 6: 4801FA ADD RDX, RDI
; 9: C9 LEAVE
; A: F8 CLC
; B: C3 RET
NILCL-USER> (defun fx-add (x y)
(declare (optimize (speed 3) (safety 0) (debug 0))
(type fixnum x y))
[OUT]: FX-ADD
CL-USER> (disassemble #'fx-add)
; disassembly for FX-ADD
; Size: 27 bytes. Origin: #x55498736 ; FX- ADD
; 36: 48D1FA SAR RDX, 1
; 39: 48D1FF SAR RDI, 1
; 3C: 4801FA ADD RDX, RDI
; 3F: 48D1E2 SHL RDX, 1
; 42: 710A JNO L0
; 44: 48D1DA RCR RDX, 1
; 47: FF1425E8070050 CALL [#x500007E8] ; #x54602570: ALLOC-SIGNED-BIGNUM-IN-RDX
; 4E: L0: C9 LEAVE
; 4F: F8 CLC
; 50: C3 RET
[OUT]: NIL
``` it looks inefficient but i think it can still be optimized better. SBCL is mediocre at best at optimizations but most people do say that. I have not heard anyone calling it as fast as or faster than C/C++. I think sbcl being mediocre has more to do with how the compiler and its optimizations are structured rather than some inherent inefficiency of common lisp.
I agree with you that lisps have a terrible UX problem. But that is slowly changing. Alive for VSCode is a nice tool if you want to try it out. I personally use neovim for my CL development. fits me well enough.
The biggest reason why someone might choose CL for performance would be the abiliy to simply hack in the compiler as it is running to make custom optimization passes, memory allocations and structuring, very minute GC and allocator control, custom instructions and ability to tell sbcl how to use them WITHOUT RELOADING SBCL.
Those are just the basics in my opinion. Performance has nothing to do with the language...except ig the start times? Beyond that its a game of memory management[which, im pretty sure you can hack too in CL]
all this low level power while still being able to write soo much abstraction that you can slap a whole ass language in CL like APL!
I would think that it takes a bit more knowledge to judge that.
You'll need to understand a bit more how to declare types to achieve better results.
Sure. ---> One could add two fixnums and get a bignum.
CL-USER 31 > (fixnump (+ MOST-POSITIVE-FIXNUM MOST-POSITIVE-FIXNUM))
NIL
Yes, Common Lisp does by default determine whether to return a bignum or fixnum.
The machine code you've shown takes care of that.
Let's see what the SBCL file compiler says:
CL-USER> (compile-file "/tmp/test.lisp")
; compiling file "/tmp/test.lisp" (written 11 JUL 2024 10:02:04 PM):
; file: /tmp/test.lisp
; in: DEFUN FX-ADD
; (DEFUN FX-ADD (X Y)
; (DECLARE (OPTIMIZE (SPEED 3) (SAFETY 0) (DEBUG 0))
; (TYPE FIXNUM X Y))
; (+ X Y))
; --> SB-IMPL::%DEFUN SB-IMPL::%DEFUN SB-INT:NAMED-LAMBDA
; ==>
; #'(SB-INT:NAMED-LAMBDA FX-ADD
; (X Y)
; (DECLARE (SB-C::TOP-LEVEL-FORM))
; (DECLARE (OPTIMIZE (SPEED 3) (SAFETY 0) (DEBUG 0))
; (TYPE FIXNUM X Y))
; (BLOCK FX-ADD (+ X Y)))
;
; note: doing signed word to integer coercion (cost 20) to "<return value>"
;
; compilation unit finished
; printed 1 note
; wrote /tmp/test.fasl
; compilation finished in 0:00:00.031
#P"/private/tmp/test.fasl"
NIL
NIL
If you want a fixnum to fixnum addition, then you need to tell the compiler that the result should be a fixnum.
(the fixnum (+ x y))
CL-USER> (compile-file "/tmp/test.lisp")
; compiling file "/tmp/test.lisp" (written 11 JUL 2024 10:03:11 PM):
; wrote /tmp/test.fasl
; compilation finished in 0:00:00.037
#P"/private/tmp/test.fasl"
NIL
NILThe "advantage" of this system was that there was no need for programmers, as there was "no code", just configuration!. This was supposed to allowed "domain experts" without programming knowledge to work with the system. However a month long training by the creator of the system was still required, as he had to explain which of the 7 boolean types you should use if you wanted to add a new column 0.o (for those who want to know, there was true/false, 0/1, yes/no, true/false/unknown, true/false rendered as a toggle, true/false rendered as a checkbox...)
This is hilarious! It takes a special kind of ignorance to come up with a solution like this.
That said, I suppose stranger things have happened.
Having worked in an org with an official in-house genius who was terribly tight with a tech-illiterate leadership and faked his way into his status, I can't really tell. Throwing people under the bus, blaming the world around them for problems created by your brittle code, shunning best practices in favor of finger-pointing... This happens in small shops more often than we'd like believe.
As the saying goes, truth is stranger than fiction. Because fiction is expected to make sense.
There's a Wikipedia page about it: https://en.wikipedia.org/wiki/Inner-platform_effect
A variant of it is: Any sufficiently complicated program contains a slow and buggy implementation of half of Lisp. That's the Greenspun's tenth rule: https://en.wikipedia.org/wiki/Greenspun%27s_tenth_rule
This applies to the kernel as well to put it bluntly and a bit ironically: eBPF, but this shouldn't be understood that I mean that eBPF is not well thought out! https://en.wikipedia.org/wiki/EBPF
I imagine everyone has invented this scheme at one point or another. It's so obvious, when you think about it!
I've encountered this pattern several times over my career. Some very smart programmer decides that for "reasons" the standard way to do something is "bad". (Usually "performance" or "bloat" are words bandied around.) They then happily architect a new system to replace the "old thing". Of course the new thing is completely undocumented (because genius programmers don't waste their time writing docs).
If you're _lucky_ the programmer then spends his whole career there maintaining the thing. If you're lucky the whole thing becomes obsolete and discarded before he retires. Hint: You're not lucky.
So what you are left with is this big ball of smoosh, with no documentation, that no-one can figure-out, much less understand. Oh he designed this before multi-core processors were a thing? Before we switched to a preemtive threaded OS? Well no, none of the code is thread-safe, and he's left the company so we need someone to "just update it".
There are reasons standard libraries exist. There are usually reasons they're a bit slower than hand-coding specific cases in assembler. There are reasons why they are "bloated" with support for lots of edge-cases. (like comments).
When some really smart person starts talking about how it's all rubbish, be afraid. Be very afraid.
That right there. Before there is a standard lib for something if there are N people coding something up there could be N! ways to do something.
If you do not know about a standard lib or it doesn't exist there will be some wild code written.
It is when that standard library shows up you should at least consider just throwing your bespoke code away. Not always but should at least be considered. I personally have replaced thousands of lines of code and modules I wrote just by switching them to some existing library. The upside is if that standard lib does not do what I want I have enough knowledge to either bend it around so it does or I can fix it up (or put my bespoke code back). I know I am not that smart, but I know enough that my code is probably brittle and probably should be thrown away.
Also watch out for some 'standard libs'. Some of them are little more than someone's hobby project and have all the exact same issues you are trying to avoid. One project I worked on some guy had written a grid control. He was charging something like 10k a year to use it. But it was just one guy and I quote "i just touch it once or twice a year and drink margaritas on the beach". It was a bug prone riddled mess we spent a non insignificant amount of time fixing. We bought another one for a onetime fee of 500 bucks and it was wildly faster and more importantly had near zero bugs and a turn around time of 1-2 days if we found one.
This is all on itself a sign that your project is fundamentally broken, but this is already covered by scenario b) incremental builds.
Even if for some reason you resist the urge of following best practices and not create your own problems, there are a myriad of techniques to limit the impact of touching a single file in your builds. Using a facade class to move your serialized JSON to an implementation detail of a class is perhaps the lowest effort one, but the textbook example would be something like a PIMPL.
The main problem with the build time of C++ projects are not the build times per se but clueless developers, who are oblivious to the problem domain, fumbling basic things and ending up creating their own problems. Once one of them stops to ask themself why is the project taking so much time to build, more often than not you find yourself a few commits away from dropping build times to a fraction of the cost. Even onboarding something like ccache requires no more than setting an environment variable.
a: You don't have source file and header file, you put everything in one file and lzz sorts it out during build.
That's the root cause of the slow build. That file is likely to be depended on by way too many other files, triggering massive rebuilds when unrelated code is modified. The headers should be as granular as possible. Breaking up the generic utils file into many specific files will help contain the damage whenever any given file is changed.
I wish it was possible to track source code dependencies at the function level.
Like, most C projects import enormous headers with gazillions of function prototypes without having a particularly measurable impact on compile times (compile times in C is quite good, in fact!)
Breaking up the headers into granular files should still help reduce the amount of instatiation that's going on at compile time provided there isn't much overlap in the headers included by the source files.
We want to reduce the set of rebuilt files to a minimum. That means separate headers so that files that use A don't need to be recompiled because B changed and A and B are defined in the same header.
Taking this logic to the extreme would lead to one file per type or function. I've read a lot of code that's structured this way and it works well. Editing lots of small files is a little annoying though. In the end it's a tradeoff.
Specifies the type of the form. In this example, it tells the CL compiler that the returned `(+ x y)` is a `fixnum`.
* (declaim (ftype (function (fixnum fixnum) fixnum) fx-add))
(FX-ADD)
* (defun fx-add (x y)
(declare (optimize (speed 3) (safety 0) (debug 0)))
(+ x y))
FX-ADD
* (disassemble 'fx-add)
; disassembly for FX-ADD
; Size: 6 bytes. Origin: #x552C81A6 ; FX-ADD
; 6: 4801FA ADD RDX, RDI
; 9: C9 LEAVE
; A: F8 CLC
; B: C3 RET
NIL