A metaprogram is built to generate files before the main program is built. We'll term runtime for this program as `gen_time`. The metaprogram's core implementation are within `gen.hpp` and `gen.cpp` in the project directory.
`gen.cpp` \`s `main()` is defined as `gen_main()` which the user will have to define once for their program. There they will dictate everything that should be generated.
**Note: The formatting shown here is not how it will look. For your desired formatting its recommended to run a pass through the files with an auto-formatter.**
The other one in the gen directory within test is the metaprogram's build specification.
Both of them build the same source file: `test.cpp`. The only differences between them is that gen need a different relative path to the include directories and defines the macro definition: `gen_time`.
This method is setup where all the metaprogram's code are the within the same files as the regular program's code.
If in your use case, decide to have exclusive separation or partial separation of the metaprogam's code from the program's code files then your build configuration would need to change to reflect that (specifically the sources).
There is no parse API for validating excution statements (possibly will add in the future, but very limited in what it can do).
This reason there isn't one: thats where the can of worms open for parsing validation.
For most metaprogramming (espcially for c/c++), expression validation is not necessary, it can be done by the compiler for the runtime program.
Most of the time, the critical complex metaprogramming conundrums are actaully producing the frame of abstractions around the expressions.
Thus its not very much a priority to add such a level of complexity to the library when there isn't a high reward or need for it.
To further this point, lets say you do have an error with an execution statment. It will either be caught by the c++ compiler when compiling the target program, or at runtime for the program.
* If its not caught by the compiler, the only downside is the error appers on the generated function. Those with knowledge of how that definition was generated know where to find the code that inlined that expression in that file for that definition.
* If its caught at runtime. The expression will be shown in a stack trace if debug symbols are enabled in the generated function body. Yet again those with knowledge of how that definition was generated know where to find the code that inlined that expression.
In both these cases will get objectively better debug information than you would normally get on most c++ compilers with complex macros or templates.
### The Data & Interface
As mentioned in [Usage](#Usage), the user is provided Code objects by calling the constructor functions to generate them or find existing matches.
The AST is managed by the library and provided the user via its interface prodedures.
However, the user may specificy memory configuration.
* The allocator definitions used are exposed to the user incase they want to dictate memory usage
* You'll find the memory handling in `init`, `gen_string_allocator`, `get_cached_string`, `make_code`, and `make_code_entries`.
* ASTs are wrapped for the user in a Code struct which essentially a warpper for a AST* type.
* Both AST and Code have member symbols but their data layout is enforced to be POD types.
* This library treats memory failures as fatal.
* The default setup assumes large definition sets may be provided to bodies so AST::Entires are dynamic arrays.
* They're allocated to arenas currently and are pretty wasteful if they go over their reserve size (its never recycled).
* Most likely will need to implement a dynamic-sized bucket allocation strategy for the entry arrays if memory is getting stressed.
* Otherwise if you are using fixed size entries and your definitions are under 128~512 entries for the body, you may be better of with a fixed-sized array.
* Strings are stored in their own set of arenas. AST constructors use cached strings for names, and content.
* The purpose is to overrite a specific file, it places its contents in a buffer to scan.
* Requests are populated using the following interface:
* add : Add code.
* remove : Remove code.
* replace: Replace code.
All three have the same parameters with exception to remove which only has SymbolInfo and Policy:
* SymbolInfo:
* File : The file the symbol resides in. Leave null to indicate to search all files.
* Marker : #define symbol that indicates a location or following signature is valid to manipulate. Leave null to indicate that the signature should only be used.
* Signature : Use a Code symbol to find a valid location to manipulate, can be further filtered with the marker. Leave null to indicate that the marker should only be used.
* Policy : Additional policy info for completing the request (empty for now)
* Code : Code to inject if adding, or replace existing code with.
Additionally if `GEN_FEATURE_EDITOR_REFACTOR` is defined, refactor( file_path, specification_path ) wil be made available.
Refactor is based of the refactor library and uses its interface.
It will on call add a request to the queue to run the refactor script on the file.
* The purpose is to grab definitions to generate metadata or generate new code from these definitions.
* Requests are populated using the add( SymbolInfo, Policy ) function. The symbol info is the same as the one used for the editor. So is the case with Policy.
The file will only be read from, no writting supported.
One great use case is for example: generating the single-header library for gencpp!
### Additional Info (Editor and Scanner)
When all requests have been populated, call process_requests().
It will provide an output of receipt data of the results when it completes.
Files may be added to the Editor and Scanner additionally with add_files( num, files ).
This is intended for when you have requests that are for multiple files.
Request queue in both Editor and Scanner are cleared once process_requests completes.
Templates also have a heavy cost to compile-times due to their recursive nature of expansion if complex code is getting generated, or if heavy type checking system is used (assertsion require expansion, etc).
Unfortunately most programming langauges opt the approach of internally processing the generated code immediately within the AST or not expose it to the user in a nice way to even introspect as a text file.
Stage metaprogramming doesn't have this problem, since its entire purpose is to create those generated files that the final program will reference instead.
This is technically what the macro preprocessor does in a basic form, however a proper metaprogram for generation is easier to deal with for more complex generation.
The drawback naturally is generation functions, at face value, are harder to grasp than something following a template pattern (for simple generation). This drawback becomes less valid the more complex the code generation becomes.
Thus a rule of thumb is if its a simple definition you can get away with just the preprocessor `#define`, or if the templates being used don't break the debugger or your compile times, this is most likely not needed.
* You enjoy actually *seeing* the generated code instead of just the error symbols or the pdb symbols.
* You value your debugging expereince, and would like to debug your metaprogram, without having to step through the debug version of the compiler (if you even can)
* Review if the upfront or incremental constructors are actually a net benefit vs just using the parse constructors.
* They exist as a artifact of learning what was possible or not possible with staged metaprogramming in C++ (the parse interface was the last to get fleshed out)
* Most likely at least Incremental could possibly be removed in favor of just using the parse constructors.
* Possible merits are ergonomics for very dynamic generation or performance reasons.
* They'll most likely stay until its evident that they are not necessary.
* Review memory handling for the AST, specifically relating to:
* Giving type asts a dedicated memory arenas.
* Giving specifier definitions a dedicated memory arenas and hashtable lookup.
* Possibly adding a dedicated block allocator for the dynamic arrays of AST::Entires.
