The C++ implementation will create expressive API calls through utilizing combination of : ▪POD structs -- structs with data members only ▪designated initializers -- initializing a struct with: { .field = the road state -- There is only one. // the .= syntax is a designated initializer used to assign POD values auto result = processCarStateTransition({ .car = Car::FastSportsCar, .startingRoadState nextRoadState = { .type = RoadType::Avenue,.speed = Speed::NORMAL}}); Ambiguity The following structs will in theory dispatch to processCarStateTransition depending on how you call them. Non-Ambiguous

14 Aside: POD Types  Standard-layout types are defined in C++11.  The 2003 Standard didn’t mention standard-layout types — it describes layout guarantees in terms of POD types.  POD stands for “Plain (There’s also a POF.)  A POD is a standard-layout class that also lacks:  private or protected non-static data members  user-declared constructors  base classes  In practice, POD classes are too restrictive Type type_; short x_, y_; }; void draw(Shape *); void rotate(Shape *, int);  Here, we’re using structs and non-member functions.  We could just as well use classes and static member functions. 97 Copyright

constructor, destructor, or copy assignment operator… This definition is essentially the same as a C++03 POD type. armv7-a System V x86 System V x86 Microsoft A Composite Type not larger than 4 bytes is returned 16 bytes x86 System V fundamental only x86-64 Microsoft 1,2,4,8 bytes, C++03 POD x86 Microsoft 1,2,4,8 bytes, C++03 POD Composite types are required to be “trivial” to get into registers!C.20: If you ___$ReturnUdt$[esp-4] mov DWORD PTR [eax], 60 mov DWORD PTR [eax+4], 0 ret 0 C H R 67 not a POD not a POD size > 4 not fundamentalCan we do something about it? ● std::chrono would have to give up encapsulation

• cascading calls ending up partly in external libraries • creating, passing and returning data structs → explicit vectorizationBackground: A particular CFD task in CODA Compute flows in a cube in every SIMD_ACCESS(x, i) + SIMD_ACCESS(y, i); }); }Extended Features: Reflections Passing around data structs 25 Olaf Krzikalla, DLR SP, 2024-09-17 a b 𝑃𝑜𝑖𝑛𝑡1 𝑥1 + 𝑥1 𝑦1 + 𝑦1 𝑃𝑜𝑖𝑛𝑡2 𝑥2 ++i) { z[i] = x[i] + y[i]; } }Extended Features: Reflections Passing around data structs ▪ Conversion to structs-of-simd with the SIMD_ACCESS library 26 Olaf Krzikalla, DLR SP, 2024-09-17 a b

better performance than the normal way to design and architect systemsThe “normal way” • Classes/structs of members of heterogeneous types: won’t ever map efficiently to the computing resources • The “abnormal” way8 Recap BeginsColumnar Representation (Scattering) • Transforming collections of structs of members of heterogeneous types into one structure with arrays of homogeneous data types (going contains the componentsHandles • Underneath they are an integer (the index of the components) • Structs/Classes ==> handleHandles auto state = State; class ItemHandle { std::size index; public:

arg<"name"> = “John"};Turning the query string into meta_structs for fields and parameters template struct meta_structs_from_query { static constexpr auto ts = parse_type_s };Why not string_view?Turning the query string into meta_structs for fields and parameters template struct meta_structs_from_query { static constexpr auto ts = parse_type_s

= get<"score">(v); assert(*std::max_element(scores.begin(),scores.end()) == 12);Duck Typing for Structs u We may have a function that doesn’t care about the type of the struct, but just the name and int main() { MyPerson p1; YourPerson p2; display_person(p1); display_person(p2); }Duck Typing for Structs u We require a template function u This means we can’t use separate compilationDuck Typing for meta_struct has those names and convertible types, it can be passed to the function.Generating Meta Structs from Compile Time StringsCompile Time Regular Expressions struct date { std::string_view year;

); } } return opts; } 1 2 3 4 5 6 7 8 9 10 11 12 *Copied from P2996 35Array of structs to struct of arrays struct point { float x; float y; }; using points = struct_of_arrays y; // }; 1 2 3 4 5 6 7 8 9 10 11 *Copied from P2996 36Array of structs to struct of arrays template struct struct_of_arrays_impl; consteval 75Full ABI textual representation of the layout You could generate a “schema” file for your structs Could be some JSON based schema Could be a pre-existing schema like Apache Avro This way we get

BOOST_METAPARSE_STRING("Hello") C++ 98: C++ 11:Boost Metaparse - Meta Functions The library uses template structs also known as meta functions to create and combine parsers. struct custom_parser { template empty_list;CTRE – LL(1) Parser The grammar rules lookup table is implemented as empty structs and function overloads struct pcre { struct backslash{}; struct hexdec_repeat {}; auto rule(backslash 115 7.834 6.709 6.274 Compile-time parsing Enabled RemovedBenchmark – Struct Parsing (1000 structs) Debug Release 0.986 1.046 1.22 1.134 15.583 15.345 0.0143 0.0142 Baseline Regular C++ Compile-time

"Primitive" - CppCon 2020 24One interface – many implementations • We can’t just • Read/write structs • Interfaces to different transports differ • Some want zero-copy • Version skew • Some want guarantees demanding application • Repeated many times in many variations • Multiple compilers • Bytes, Structs, Accessors, out-of-order access: about 4 ns • Getters and setters about 9 ns Stroustrup - "Primitive"