CPP / C++ - Videos and Conferences

• Engineering Distinguished Speaker Series: Howard Hinnant - YouTube
  • "Howard Hinnant spoke at Bloomberg presenting everything you need to know about move semantics."

Notes:

"Special member functions are those member functions that the compiler can be asked to automatically generate code for."

In C++ 98, there were 4 special member functions that the compiler can generate if not specified, namely:

In C++11, there 6 are member functions that the compiler can generate:

A special member function can be:

• not declared
• implicitly declared (The compiler can be asked to generate the function automatically)
  • delete annotation C++11
  • default annotation C++11

C++ => Declaration X Definition (aka implementation)

• Declaration: only method signature, class signature … without any definition. Generally in *.hpp files (header files).
• Definition: a full definition => Function Signature + Body => Generally in *.cpp files.

User-declared:

struct X
{
  /* Default Constructor: User declared => The definition or implementation 
   * is elsewhere, generally in *.cpp file. 
   */
  X() {} 

  X();           // User declared 
  X() = default; // User declared
  X() = delete;  // User declared, any attempt to use a deleted member
                 // functions yields a compile-time error.
};

Default constructor and overload resolution (1):

• If the template has zero arguments, the preferred constructor will be the CTOR2 since according to C++ standard, the non-templated function will be preferred in an overload resolution.
  • TL;DR => Function CTOR2 gets called.

struct X
{
   // ===> CTOR1 (Variadic constructor)
   // Universal reference 
   // (It can be either, L-value reference or R-value reference.)   
   template <typename ... Args> 
    X(Args&& ... args); // Implementation is elsewhere 

   // ====> CTOR2 (If the template has zero args, this function gets called.)
   X() = default; 
};

Default constructor and overload resolution (2):

• deleted members pariticipate in overload resolution.
• members not-declared DO NOT pariticipate in overload resolution.
• Any attempt to call the default constructor X() (CTOR2) will result in a compile-time error.

struct X
{
   // ===> CTOR1 (Variadic constructor)
   // Universal reference 
   // (It can be either, L-value reference or R-value reference.)   
   template <typename ... Args> 
    X(Args&& ... args); // Implementation is elsewhere 

   // ====> CTOR2 (If the template has zero args, this function gets called.)
   X() = delete; 
};

Under what circunstances are special members implicitly provided by the compiler.

• A user-declared destructor will inhibit the implicit declaration of the move members.
• (Row D) - The implicitly defaulted copy members are deprecated.
• (Row D) RULE OF THREE: If any of destructor, copy-ctor or copy-assignment is declared, all should be declared, even if not necessary, specially in classes managing resources or heap-memory (free store).
• (Row F) A user declared copy-constructor inhibit move members and the copy-assignment operator is deprecated. It should defined and declared explicitly. (RULE OF THREE)
• (Row H) If only the move constructor is declared, the compiler sets the copy constructor and copy assignment as deleted and the constructor is not declared. The underlying class becomes a move-only class.
  • Example: std::unique_ptr.

Compiler Implicitly Declares for C++11

• Note: The defaulted functions are only generated if they are used. It means, invoked in some part of the code.

Compiler Implicitly Declares for C++98/03

The function std::move, just casts an L-value (X&) to R-value (generally temporary objects from which it is not possible to take an address) (X&&). It has to use, to force the invocation of move constructor or move assignment operator for transfering resource ownership.

All STL containers have move constructors and move assignment operator by default which makes efficient to return them by value from functions.

Can I define one special member function in terms of another?

• YES. It turns out that it is possible.
• However, the lecturer advises to NOT DO SO. Advice: "Give each special member function the tender loving care it deserves." Specially move semantics which can enhance performance significantly.
• The entire point of move semantics is to boost performance.

Copy and Swap Idiom

• Note: This trick is neat, however it has some performance cost. The code can be optimized by writing both copy-assignment and move-assignment individually.
• Benefit: Strong exception safety

class X
{
 std::vector<int> m_v;
 public: 
 // Implements both copy and move assignment operator. 
 X& operator=(X x){
    m_v.swap(x.m_v);
    return *this;
 }
};

Rule of Zero

• If the class only uses STL containers as member functions (not pointers or unrwapped resources such as socket handlers), most of the time, it is not necessary to define the special member functions explicitly and the default generated member functions already have optimal performance.

class X
{
 std::vector<int> m_v;
 public: 
 // ===> Everything declared explicitly 
 // as explicit is always better than implicit. 
 //----------------------------------

 /* Default ctor */
 X() = default; 
 /* Default dtor */
~X() = default; 
 /* Default copy-ctor */
 X(const X&) = default; 
 /* Default move-ctor */
 X(X&&) = default;
 /* Copy assignment operator */
 X& operator=(const X&)  = default;
 /* Move assignment operator */
 X& operator=(const X&&) = default;
};

Trick for strong exception sfaety on assignment

Note: It is an alternative to copy-and-swap diim.

template <class C>
C& strong_assign(C& dest, C src){
  std::swap(dest, src);
  return dest;
}

Usage:

// Instead of assignment 
x = y; 

// Use - strong_assign where 
// strong exception safety is needed. 
// Gets the best of both worlds.
strong_assign(x, y);

In a hurry?

If there is not time think carefully, the copy members should be deleted. It disables all the special members, except the destructor which is implicitly declared. Any attempt to use any of disabled members functions results in a compile-time error.

class X
{
 public:
    X(const X& X) = delete;
    X& operator=(const X& X) = delete;
};

What is the state of a moved-from object?

• The state from a moved-from object is often unspecified. It is not the same as undefined behavior, just the object's state is not known.

What is a precondition?

• A requirement in the function specification which restricts the state of the object prior to the call. For instance,
  • std::vector<T>::pop_back()
  • Requires the PRECONDITION: empty() shall be false before changing the state of the vector.

Summary

• Known when the compiler is defaulting or deleting special member for you, and what defaulted memebrs will do.
• Always define or delete special members whn the compiler's implicit action is not corret.
• Give tender lovign care to each of 6 special member functions, even if the result is to let the compiler handle it.

Dive into C++11

• Dive into C++11 - <1> - Arkanoid clone in 160~ lines of code (SFML 2.1)
• Dive into C++11 - <2> - Frametime, FPS, constexpr, uniform initialization
• Dive into C++11 - <3> - Automatic lifetime, pointers, dynamic allocation
• Dive into C++11 - <4> - Smart pointers
• Dive into C++11 - <5> - Game entity management basics
  • Entity component system.

Dive into C++14

• Dive into C++14 - <1> - Introduction to C++14 core language features
  • Note 1: Contexpr => Can have loops, for-loops and doesn't need to be recursive as before. It is useful for computing look-up tables, hash, CRC32 and checksums at compile-time.
  • Note 2: Return type deduction => Return type can be omitted. However, it can be make the code harder to read.
  • Note 3: Variable templates. Allows computing numerical constats at compile-time such as PI with some precision.
  • Note 4:
    • Lambda functions can have auto argument or argument type deduction.
    • Lambda functions can have variadic parameter pack.
• Dive into C++14 - <2> - `for_each_argument` explained and expanded
• Dive into C++14 - <3> - Generic "unique resource" wrapper

• CppCon 2017: Bryce Adelstein Lelbach “C++17 Features (part 1 of 2)” - YouTube
  • The feature set for the C++17 release is set, and the release of the standard is just around the corner. In this session, we'll discuss all the new C++ features in C++17 and how they'll change the way we write C++ software. We'll explore the new standard in breath, not width, covering a cornucopia of core language and library features and fixes:
  • Language Changes (part 1):
    • Structured bindings
    • Selection statements with initializers
    • Compile-time conditional statments
    • Fold expressions
    • Class template deduction
    • auto non-type template parameters
    • inline variables
    • constexpr lambdas
    • Unary static_assert
    • Guaranteed copy elision
    • Nested namespace definitions
    • Preprocessor predicate for header testing
  • Library Changes (part 2):
    • string_view
    • optional
    • variant
    • any
    • Parallel algorithms
    • Filesystem support
    • Polymorphic allocators and memory resources
    • Aligned new
    • Improved insertion and splicing for associative containers
    • Math special functions
    • Variable templates for metafunctions
    • Boolean logic metafunctions

• Bjarne Stroustrup - The Essence of C++
  • "Bjarne Stroustrup, creator and developer of C++, delivers his talk entitled, The Essence of C++. Stroustrup has held distinguished posts at Texas A&M University and spent significant time in the Computer Science Departments of Cambridge, Columbia and Princeton. C++ is the one of the world's most widely used technology languages and it has influenced newer languages such as C# and Java as well as older languages."
    • Applications on System Programming Domains:
      • Device drivers
      • Network drivers
      • Embedded Systems
      • Suitable for hard and soft real time systems.
    • C++ Origin: C + Simula (First Object Oriented language)
    • Main Ideas:
      • Direct map to hardware
      • Zero-overhead abstraction
      • Much of inspiration comes from operating systems.
      • Backward compatibility: Avoid breaking old code.
    • Key strengths:
      • A language for building lightweight abstractions.
      • Software infrastructure.
      • Resource-constrained application. (May be embedded systems)
      • Resource Management

• CppCon 2017: Michael Spencer “My Little Object File: How Linkers Implement C++” - YouTube
  • "Ever wonder how the linker turns your compiled C++ code into an executable file? Why the One Definition Rule exists? Or why your debug builds are so large? In this talk we'll take a deep dive and follow the story of our three adventurers, ELF, MachO, and COFF as they make their way out of Objectville carrying C++ translation units on their backs as they venture to become executables. We'll see as they make their way through the tangled forests of name mangling, climb the cliffs of thread local storage, and wade through the bogs of debug info. We'll see how they mostly follow the same path, but each approach the journey in their own way. We'll also see that becoming an executable is not quite the end of their journey, as the dynamic linker awaits to bring them to yet a higher plane of existence as complete C++ programs running on a machine."

• Modern C++ in Embedded Systems - YouTube
  • "For nearly 35 years I have been working with small processors and there has always been deep divides between practitioners of languages. When writing assembly we ridiculed those using C and when I spent years microcoding we scoffed at everyone. However, nearly all groups continue to wag their heads at the shameful C++ programmers attempting to twist their tools toward the small. Recent language developments have made C++ the obvious choice for many embedded projects; nevertheless, the toxic environment extends past reddit roasts into poor vendor support of tools and nearly obstructionist chip manufacturers. This session will use a bare-metal project started in 2018 for a Ciere client as a case study while we work through the decision process for using C++, the acrobatics required to support the language, recent language features that enable goals of size, speed, and expressiveness, and useful libraries. While the examples will be based on a concrete project, the extracted lessons-learned should be applicable to many embedded projects (bare-metal and small-OS). Attendees will walk away with motivations to use C++ in embedded projects, hints and tips to making tools work, and a sampling of language features and idioms that improve the quality of a final product."

See: https://arobenko.gitbooks.io/bare_metal_cpp/content/

What is shown:

• Case presented: An ARM-based embedded processor is used for controlling motors and a hydraulic machine.
• Processor used: ARM Cortex RC4 connected to a FPGA (Filed Programmable Gate Array).
• Peripherals mentioned: I2C, SPI
• Tooling:
  • GCC toolchain for ARM
  • CMAKE for build automation.
  • Linker script
• State Machine => Boost.SML (https://youtu.be/1l2g2dAobXA?t=4335)
  • Typical state machines used: Mealy's state machine.
  • https://github.com/boost-experimental/sml
• LADON - Real time object oriented modelling from 90's (prodecessor for UML Real time extensions) => Distributed State Machines.
  • Allows to connect multiple state machines.
• Metaprogramming techniques used:
  • tag dispatching
  • CRTP => Curious Recurring Template Pattern for emulating dynamic polymorphism (runtime poly.) with static polymorphism at compile time.
  • Mixins
  • Variant Polymorphism
  • Type traits (metafunctions) => struct which takes list of types as arguments peforming computations on types.
  • Type synonyms with 'using' keyword
• Common compiler options disabled on code for embedded systems:
  • Disable RTTI (-fno-rtti)
  • Disable exceptions (-fno-exceptions)
• How can C++ simplify the project?
  • More declarative.
  • More correctness and bugs caught at compile-time.
  • Template metaprogramming is crucial for developing lightweight abstraction and reducing run-time overhead.

• CppCon 2015: Herb Sutter "Writing Good C++14… By Default"
  • Total time: 1:29:05
  • "Modern C++ is clean, safe, and fast. It continues to deliver better and simpler features than were previously available. How can we help most C++ programmers get the improved features by default, so that our code is better by upgrading to take full advantage of modern C++?. This talk continues from Bjarne Stroustrup’s Monday keynote to describe how the open C++ core guidelines project is the cornerstone of a broader effort to promote modern C++. Using the same cross-platform effort Stroustrup described, this talk shows how to enable programmers write production-quality C++ code that is, among other benefits, type-safe and memory-safe by default – free of most classes of type errors, bounds errors, and leak/dangling errors – and still exemplary, efficient, and fully modern C++."

Core Suppport Guidelines ISO CPP:

• http://github.com/isocpp/CppCoreGuidelines

Core Support Guideline Microsft GSL:

• http://github.com/microsft/gsl

Old X New, modern C++

Moment: At 1:55

Old C++ 98:

• Problem: code hard to reactor, hard to make changes as it can cause memory leaks.
• Any changes in the code requires verification for memory leaks and dangling pointers.
• A great deal of security vulnerabilities are related to memory leaks.

circle* p = new circle(42);
vector<shape*> v = load_shapes();

for(vector<shape*>::iterator i = v.begin(); i != v.end(); ++i)
{
  if(*i && **i == *p)
     cout << **i << "is a match\n";
}
//.. later, possibly elsehwere 
for(vector<shape*>::iterator i = v.begin(); i != v.end(); ++i)
{
   delete *i;
}

delete p;

Modern C++: (>= C++11)

• Advantages:
  • Easier to read and more memory-safe.
  • Modifications do not require following pointers around on every possible return path and checking for memory leaks or double deleting.

// Smart pointers for memory safety  
auto p = std::make_shared<circle>(42);

// Usage of auto keyword for type deduction 
auto v = load_shapes();

// For-loop uses the new notation instead of iterators.
for(auto& s: v)
{
  if(s && *s  == *p)
    cout << *s << "is a match\n";
}

GSL - Guideline Support Library - Safety - At 6:06

Core Guidelines - Safe replacements (GSL)

• Type safety + concepts + module s
• Bound safety + contracts
• Lifetime safety

Initial Target: Type & Memory Safety

• Traditional definitin:
  • = type-safe
  • (+) bound-safe:
    • Don't go outside the bounds of an allocation, for instance, go outside the bounds of an array.
  • (+) lifetime-safe
    • Don't access an invalid memory.
• Examples:
  • Type: Avoid unions, use variants
  • Bounds: Avoid pointer arithmetiuc, use array_view
  • Lifetime:
    • Don't leak (do not forget to delete)
    • Don't corrupt (double-delete)
    • Don't dangle:
      • use a pointer to an already deleted object
      • return a reference to locally allocated value
  • Future: Concurrency (thread-safe), security …

Target Gurantee: 8:49

• Goal is not to provide verifed, whole-program guarantees of safety.
• Goal is enable type and memory safety by construction, for as much of your program code as possible. This type and memory safe can be enforced at compile time via static language subset restrictions + at runtime by validations/eforcement (fail-fast, or configurable)

C++ code compile in the safe subset is never the root cause of type/memory safety errors, except where explictly annotated as unsafe.

Safety profiles

• A profile is:
  • A cohesive set of deterministic portable suset rules
  • designed to achieve a specific guaranteee
• Benefits of decomposed profiles:
  • Articulates what guarantees you get for what effort.
  • Avoids monolith "safe/unsafe" when opting in/out
  • Extensible to future safety profile (e.g, security, concurrency, arithmetic, noexcept, noalloc, …)
  • Enables incremental development/delivery

Type safety overview - 11:29

• GSL Types:
  • byte: Raw memory is not a char (character)
  • variat<..Ts> Contains one object at a time (tagged union)

Rules:

1. Don't use reinterpret_cast
2. Don't use static_cast downcast, use dynamic casts instead.
3. Don't use const_cast to cast away consts
4. Don't use C-tstyle (T) expression casts that would perform reinterpret_cast, static_cast downcast, or const_cast.
5. Don't use a loca variable before it has been initialized.
6. Alaways initialize a member variable
7. Avoid accessing members of raw unions. Prefere variants instead.
8. Avoid reading from varargs or passing varargs arguments. Prefere variadic.

Also: safe math -> Separate profile.

Bounds safety overview - at 13:43

Rules:

1. Do'nt use pointer arithmetic. Use array_view instead.
2. Only index into arrays using constant expresison.
3. Don't use array-to-ppointer decay.
4. Don't use std::functions and types that are bounds-checked.

Applying a profile: Explicit opt-out - at 20:57

Lifetime Profile - at 24:34

• Delete every heap object (no leaks)
• Delete every heap object only once. (no memory corruption)
• Don't dereference * to a delete object (no dangling)

PSA: Pointers are not evil. - at 28:32

• Smart pointers are good - they encapsulate ownership
  • Observation: They should only be used where there is ownership: with heap-allocated objects or polymorphic objects of some class hierarchy that oftens need dynamic allocation.
• Raw T* and T& are good - we want to maintain the efficiency of "just an address", specially on the stack (locals, parameters, return values)
  • Observation: Better used as function parameters and for parameter passing, even for smart pointers when there is no ownership.

Lifetime Safety Overview - at 29:26

GSL types, aliases, concepts:

• Indirect concept:
  • Owner (can'ty dangle): owner<T>, container, smart pointers, …
  • Pointer(could dangle): *, &, array_view,/string_view, ranges, …, not_null<T>, Warps any indirection and enforces non-null.
• Owner<>: Alias, ABI-compatible, building block for smart ptrs, container…
  • Main owner<T*>

Rules:

1. Prefer to allocate heap object using make_unique, *make_shared_ or containers.
2. Otherwise, use owner<> for source/layout compatibility with old code. Each non-null owner<> must be deleted exactly once, or moved.
3. Never dereference a null or invalid pointer
4. Never allow an invalid pointer to escape a function.

• CppCon 2016: Herb Sutter "Leak-Freedom in C++… By Default."
  • Time: 1:39:24
  • "Lifetime safety means writing code that, by construction, is guaranteed to eliminate two things: (a) use of null/dangling pointers (including pointerlike things such as references, iterators, views, and ranges), and (b) leaks (including the rare 1% case where we’re tempted to admit the possibility of an ownership cycle or need to support lock-free concurrent data structures). Last year, my CppCon 2015 talk “Writing Good C++14… By Default” focused on (a), null/dangling, because it's the more difficult and usually more serious problem. I gave an overview of a new approach of using static analysis rules to eliminate use of null and dangling in C++. That work continues and we’re in the process of writing down the formal rules for the approach that I showed last year. This year, the focus will be on (b), leaks: The talk aims to begin with a set of simple rules, the “5-minute talk” to demonstrate that a handful of rules can be taught broadly to programmers of all levels, and results in code that is clean and free of leak bugs by construction. But, since we’ll still have 85 minutes left, we can use the time to spelunk through a series of “Appendix” code examples, in which we'll demonstrate "why and how" to apply those rules to a series of increasingly complex/difficult situations, and that are aimed at increasingly advanced and “clever” (note: not always a good thing) programs and programmers. We’ll address questions such as: How should we represent Pimpl types? How should we represent trees – what should the child and parent pointer types be, and (when) should they be unique and when shared? How should we deal with “intra-module” or “encapsulated” cycles when you control all the objects in the cycle, such as all the nodes within a Graph? And what about “inter-module” or “compositional” cycles when you don’t know in advance about all the objects that could be in the cycle, such as when combining libraries written by different people in a way that may or may not respect proper layering (notoriously, using callbacks can violate layering)? The answers focus on cases where we have solid guidance, and then move toward some more experimental approaches for potentially addressing the ~1% of cases that aren’t yet well covered by unique_ptr, shared_ptr, and weak_ptr."

• CppCon 2018: Victor Ciura "Regular Types and Why Do I Care ?"
  • Description: "“Regular” is not exactly a new concept (pun intended). If we reflect back on STL and its design principles, as best described by Alexander Stepanov in his 1998 "Fundamentals of Generic Programming” paper or his lecture on this topic, from 2002, we see that regular types naturally appear as necessary foundational concepts in programming. Why do we need to bother with such taxonomies ? Well, the STL now informally assumes such properties about the types it deals with and imposes such conceptual requirements for its data structures and algorithms to work properly. The new Concepts Lite proposal (hopefully part of C++20) is based on precisely defined foundational concepts such as Semiregular, Regular, EqualityComparable, DefaultConstructible, LessThanComparable (strict weak ordering), etc. Formal specification of concepts is an ongoing effort in the ISO C++ Committee and these STL library concepts requirements are being refined as part of Ranges TS proposal (＜ experimental/ranges/concepts＞). Recent STL additions such as string_view, tuple, reference_wrapper, as well as new incoming types for C++20 like std::span raise new questions regarding values types, reference types and non-owning “borrow” types. Designing and implementing regular types is crucial in everyday programing, not just library design. Properly constraining types and function prototypes will result in intuitive usage; conversely, breaking subtle contracts for functions and algorithms will result in unexpected behavior for the caller. This talk will explore the relation between Regular types (and other concepts) and STL containers & algorithms with examples, common pitfalls and guidance."

Coverage:

• Values
• Objects (Not object-oriented programming)
• Concepts
• Ordering Relations
• Requirements

Types: (Tintus Winters - Modern C++ API Design)

• Type Properties
  • What properties can we use to describe types?
• Type Families
  • What combinatons of type properties make useful / good type design?

Books References and Papers

• Elements of Programming - Alexander Stepanov and Paul Macjones.
  1. Foundations
  2. Transformations and their orbits
  3. Associative operations
  4. Linear ordering
  5. Order Algebraic Structures
  6. Iterators
  7. Coordinate Structures
  8. Coordinates with Mutable Successors
  9. Copying
  10. Rearrangements
  11. Partitions and merging
  12. Compose Objects
• FM2GP From Mathematics to Generic Programming - Alexander Stepanov and Daniel E. Rose
  • Egyptian multiplication - 1900 - 1650 BC
  • Ancient Greek Numebr Theory
  • Prime Numbers
  • Euclid's GCD Algorithm
  • Abstractions in mathematics
  • Deriving Generic Algorithms
  • Algebraic Structures
  • Programming Concepts
  • Permutations Algorithms
  • Cryptology (RSA) - 1977 AD
• Fundamentals of Generic Programming - James C. Dehert and Alexander Stepanov 1998
  • http://stepanovpapers.com/DeSt98.pdf
  • Quote: "Generic Programming depends on the decomposition of programs into components which may be developed separately and combined arbitrarly, subject only to well-defined interfaces".
  • Quote: "Among the interfaces of interest, the most pervasively and unconsciously used, are the fundamental operators common to al C++ built-in types, as extended to user-defined types, e.g copy constructors, assignment, and equality."

Mathematic Really Does Matter!!!

• Stepanov on GCD - One simple algorithm refined andimrpoved for over 2,500 years while advancing human udnerstanding of mathematics.
• http://www.youtube.com/watch?v=fanm5y00joc
  • "This talk by Alexander Stepanov was presented and recorded at SmartFriends U, September 27, 2003. Originally presented as the 1999 Arthur Schoffstall Lecture in Computer Science and Computer Engineering at the Rensselaer Polytechnic Institute. Introduction by Jon Kalb. Uploaded by permission of SmartFriends. Thanks to Scott Boyd and everyone involved for their production work".

Terminology

• Datum:
  • Finite sequence of zeros or ones, 0s or 1s.
• Value Type:
  • A value type is a correspondence between a species (abstract / concrete) and a set of such datums.
• Value:
  • A datum alongside its interpretation: an integer in 32-bit 2's two's complement, big endian (byte alignment).
    • min value: -(2^{32 - 1} - 1), max value: 2^{32 - 1};
  • A value cannot change
• Valye Type & Equality
  • Lemma 1: If a value is uniquely represented, equality implies representational equality. (Note: representational equality, structural equality is the opposite of OOP's identity equality where two objects are equal if they have the same memory adress, same identity)
  • Lemma 2: If a value type is not ambiguous representational equality implies equality.
• Object:
  • An object is a representation of concrete entity as a value in computer memory (address & length).
  • An object has a state that is a value of some value type.
  • The state of an object can change.
• Type:
  • A type is a set of values with smae interpretation functin and operations on these values.
• Concept (Generic Programming / Template Metaprogramming Concepts)
  • A concept is collection of similar types.

STL and Its Design Principles (Stepnaov talk)

• Parper PDF:
  • http:/stepanovpapers.com/stl.pdf
• Talk at AdobeSystems:
  • https://www.youtube.com/watch?v=COuHLky7E2Q
• Provides design guidance on extending the STL (Standard Template Library).
• Fundamental Principles:
  • Sistematically identifying and organizing useful algorithms and data structures.
  • Finding the most general representation of algorithms.
  • Using whole-part value semantics for data structures.
  • Using abstractions of addresses (iterators) as the interface between algorithms and data structures. (side-note: iterator - generalization of pointers)
  • Algorithms are associated with a set of common properties:
    • Eg. { +, *, min, max} => Associative operations
    • => Reorder operands
    • => parallelize + reduction (std::accumulate)
    • Natrual extension of 4,000 years of mathematics
    • existing generic algorithms behind while() or for() loop.
  • STL data structures (Note: Based on value semantics)
    • STL data structures extend the semantics of C structures
    • Two objects never intersec (they are separate entities)
    • thow objects have separate lifetims
  • STL data structures have whole-part semantics
    • Copy of whole, copies the parts
    • whe the whole is destroyed, all parts are destroyed
    • two things are equal when they have the same number of parts and their corresponding parts are equal.
  • Generic Programming Drawbacks:
    • abstraction penalty
    • implementation in the interface
    • early binding
    • horrible error message (nor formal specification of interfaces, yet) [CONCEPTS]
    • duck typing
    • algorithm could work on some data types, but fail to work/compile on some ohter new data structures (different iterator category, no copy semantics and so on.)
    • It is necessary to fully specify requirements on algorithm types.
    • Additional: Hard to communicate the requirements for type arguments.
  • Named Requirements - Examples From STL (Generic Programming concepts) - http://en.cppreference.com/w/cpp/named_req
    • Constructor:
      • DefautConstructible, MoveConstructible, CopyConstructible
    • Destructor:
      • Destructible
    • Equality:
      • EqualityComparable, LessThanComparable
    • Predicate:
      • Predicate, BinaryPredicate
    • Function:
      • FunctionObject
    • Container:
      • Container, SequenceContainer, ContiguousContainer, Associative Container
    • Iterator:
      • InputIterator, OutputIterator
      • ForwardIterator, BidirectionalIterator, RandomAccessIterator
  • What is the point of this?
    • Understand how the STL works and how it was designed.
    • Using STL algorithms and data structures in a better and safer way.
    • Design and expose the API vocabulary types (interfaces, APIs - Application Programming Interfaces)
    • Makes easier to read and understand the STL documentation.

Example: "GP - Concepts" and type expectations of STL Algorithm std::sort

• Compare Concept: Compare << BinaryPredicate << Predicate << FunctionObject << Callable
• Why is this one special? About 50 STL facilities(algorithms and data structures) expects some Compare type.

For instance:

template<class RandomIt, class Compare>
constexpr void sort(RandomIt first, RandomIt last, Compare comp);

Compare Concept: Ordering relationship needed.

Note:

• The symbol (∀) universal quantifier means "for all" stuffs which meets some mathematical requirement.
• The symbol (=>) arrow means "leas to".
• See:
  • Partial ordering: https://en.wikipedia/wiki/Partially_ordered_set
  • Strict weak ordering: https://en.wikipedia/wiki/Weak_ordering#Stric_weak_orderings

Code example:

#include <iostream> 
#include <vector> 
#include <algorithm> 

struct Point { 
  int x; int y; 
};
std::ostream& operator<<(std::ostream& os, Point const& p) {
    return os << "Point{" << " x = " << p.x << " ; " << " y = " << p.y << "}";
}

template<class Element>
std::ostream& 
operator<<(std::ostream& os, std::vector<Element> const& vec) {
    os << "Vector[" << vec.size() << "]{ ";
    for(const auto& x: vec) os << x << " ; " ;
    os << " }";
    return os;
}   

int main(){
   std::vector<Point> points{ {3, 4}, {10, 6}, {8, 9}, {10, 5}, {0, 0}, {-10, -4}};
   std::cout << " => BEFORE: points = " << points << std::endl;
   std::sort(points.begin(), points.end(), [](Point const& p1, Point const& p2)
   {
      if(p1.x < p2.x) return true;
      if(p2.x < p1.x) return false;
      return p1.y < p2.y;
   });
   std::cout << " => AFTER: points = " << points << std::endl;    
   return 0;
}

Output:

=> BEFORE: points = Vector[6]{ Point{ x = 3 ;  y = 4} ; Point{ x = 10 ;  y = 6} ; Point{ x = 8 ;  y = 9} ; Point{ x = 10 ;  y = 5} ; Point{ x = 0 ;  y = 0} ; Point{ x = -10 ;  y = -4} ;  }
=> AFTER: points = Vector[6]{ Point{ x = -10 ;  y = -4} ; Point{ x = 0 ;  y = 0} ; Point{ x = 3 ;  y = 4} ; Point{ x = 8 ;  y = 9} ; Point{ x = 10 ;  y = 5} ; Point{ x = 10 ;  y = 6} ;  }

Semiregular Concept

• A type which conforms to semiregular concept is:
  • DefaultConstructible
  • MoveConstructible
  • CopyConstructible
  • CopyAssignable
  • Swappable
  • Destructible

Regular Concept (Stepanov's Regular)

• Regular = SemiRegular + EqualityComparable
• STL assumes equality is always defined (at leas, equivalence relation)
• STL algorithms assumes REGULAR data structures.
• Stepanov's concept of equality: "Two objects are equal if their corresponding parts are equal (applied recursively), including remote parts (but not comparing their addresses), excluding inessential components, and excluding which identity related object." – http://stepanovpapers.com/DeSt98.pdf

Equality Comparable

• The type must work the operator== and the result shoudl have stadnard semantics.
• See:
  • https://en.wikipedia.com/wiki/Equivalence_relation
  • https://en.cppreference.com/w/cpp/named_req/EqualityComparable

C++20 - Three-way comparison - spaceship operator

• Operator <=> - http://wg21.link/p0515

(a <=> b) <  0   if  a < b 
(a <=> b) >  0   if  a > b
(a <=> b) == 0   if a and b are equal/equivalent

Monadic interface proposal (m-word)

• map() / and_then() / or_else()
• https://wg21.tartanllama.xyz/monadic-optional

Borrowed Types like std::string_view

• Borrow types are essentially "borrowed" references to existing objects.
• They lack ownership
• they are short-lived
• they generally cand do without an assignment-operator
• they generally apperar only in function parameter list
• they gnerally cannot be stored in data structured or be returned safely from function (no ownership semantics)
• https://quuxplusone.github.io/blog/2018/03/27/string-view-is-a-borrowed-type

• Better Code: Runtime Polymorphism - Sean Parent - YouTube
  • Description: "This talk explains why (and how) to implement polymorphism without inheritance in C++. The talk contains many C++ tips and techniques, including many new features from C++11. During the course of that talk a key feature from Photoshop will be demonstrated and implemented."
  • Notes: This talk is about how to preserve value semantics and polymorphism without using inheritance in the client code. It is done by using the type erasure design pattern. Actually, the solution presented do use inheritance, but inside the object_t class.
  • Benefits of the presented type erasure implementation:
    • Greater interoperability between software components as classes doesn't need to inherit froms same base class, the only requirement is to satisfy the constructor template constraints.
    • Alows value semantics to cohexist with polymorphism with minimal performance penalty. It combines the benefits of dynamic and static polymorphism.
    • A factory function can return a value rather than a pointer or smart pointer and still preserve the value semantics.
    • Non-intrusive for client code. A client code doesn't need to be aware of polymorphism. For instance, a client code double describeArea(const Shape& sh) cannot be used directly with objects returned by a factory function, for instance std::unique_ptr<Shape>. By using type erasure, this code doesn't need to modified as the factory function can returns a value instead of a pointer and still preserve the polymorphism.
      • Less error prone
      • Client code doesn't need heap allocation. (Actually, the heap allocation is hidden in the type erasure wrapper)
      • The client code doesn't need to worry about object ownership or lifetime.
      • Exception safe
      • Thread safe
    • Referenced Book: Stepanov, Alexander and Paul McJones - Elements of Programming - Addison Wesley - Professional 2009.
  • Key Moments
    • https://youtu.be/QGcVXgEVMJg?t=2032
      • The class structure is composed by concept class (interface), model classes inheriting the concept and the outter class that hides the inheritance and wraps any argument which statifies the template constraints.
    • https://youtu.be/QGcVXgEVMJg?t=2072
      • A virtual copy constructor is used in the concept class for allowing the wrapper class (outer class) object_t to preserve copy semantics by performing deep copying, instead of shallow copying.
    • https://youtu.be/QGcVXgEVMJg?t=2136 and https://youtu.be/QGcVXgEVMJg?t=2156
      • Templates used at constructor and the model.

• C++Now 2018: Louis Dionne “Runtime Polymorphism: Back to the Basics” - YouTube
  • Description: "C++ solves the problem of runtime polymorphism in a very specific way. It does so through inheritance, by having all classes that will be used polymorphically inherit from the same base class, and then using a table of function pointers (the virtual table) to perform dynamic dispatch when a method is called. Polymorphic objects are then accessed through pointers to their base class, which encourages storing objects on the heap and accessing them via pointers. This is both inconvenient and inefficient when compared to traditional value semantics. As Sean Parent said: Inheritance is the base class of evil. It turns out that this is only one of many possible designs, each of which has different tradeoffs and characteristics. This talk will explore the design space for runtime polymorphism in C++, and in particular will introduce a policy-based approach to solving the problem. We will see how this approach enables runtime polymorphism with stack-allocated storage, heap-allocated storage, shared storage, no storage at all (reference semantics), and more. We will also see how we can get fine-grained control over the dispatch mechanism to beat the performance of classic virtual tables in some cases. The examples will be based on a real implementation in the Dyno library [1], but the principles are independent from the library."
    • Problem: C++ sub-typing polymorphism inevitably requires pointers to objects allocated on the heap. However by using pointers, the advantages of value semantics, which C++ is built on top, are lost. For instance, pointers doesn't play well with C++ algorithms and functions expecting value or reference parameters. Another issue is that pointers to objects allocated on the heap raises questions about memory ownership such as who should delete the pointer.
      • Ideas => some possible approaches:
        • Use static polymorphism: template functions.
        • Use a proxy object allocated on the stack. The object inherits the base class and takes a pointer to base class as argument. Then the proxy object can forward any method call or message to the wrapped heap object.
        • Deal with it and accept as it is for performance reasons.

• Bjarne Stroustrup - Object Oriented Programming without Inheritance - ECOOP 2015 - YouTube
  • Explanative talk about generic programming (templates) paradigm as an alternative to object oriented programming inheritance.
  • Selected Moments:
    • Deep Roots in earliste C++
    • Resources and resources management
    • Move Semantics
    • Generic Programming in C++
    • An Example of generic interface
    • C++ 'concepts' (Generic programming.)

• C++Now 2018: Odin Holmes “C++ Mixins: Customization Through Compile Time Composition” - YouTube
  • " Working in the embedded domain where 'special cases' are the norm I often find myself fighting with customization capabilities and 'canned' (non-customizable) abstractions. std::string often has a small buffer for small string optimization, std::function has something similar, why can't I set the size of that buffer, or more radically why can't I just stick that same buffer in a std::vector for a 'small vector optimization'. While we're at it why can't I take .at() out of std::vectors public interface and call some user-defined function on out of memory, maybe I turned off exceptions but still want to handle errors properly. Maybe I want a std::deque interface but have a fixed sized ring buffer under the hood. Those following the SG14 will notice I am ticking off proposals, the problem with these proposals is that the mostly follow a common pattern: "I need X without Y" or "I need X to have Y" and there are many many combinations of X and Y. Why do I have to pay for atomic ref counting in a shared_pointer in a single threaded use case? We could go on all day. In this talk we will explore the feasibility of building classes from composable pieces, some concrete proof of concepts as well as the library infrastructure needed for this task."

• CppCon 2018: Mateusz Pusz “Effective replacement of dynamic polymorphism with std::variant” - YouTube
  • This short talk presents how easy it is to replace some cases of dynamic polymorphism with std::variant. During the lecture, we will analyze and compare 2 implementations of the same simple Finite State Machine. It turns up that variant-based code is not only much faster but also it gives us the opportunity to define our interfaces and program flow much better. The talk will end up with the discussion of pros and cons of each approach and will try to give guidelines on when to use them.
  • Domain: Capital Markets - HFT - High Frequency Trading that requires super fast low latency applications.

Latency

• Time required to perform some action or to produce some result.
• Measured in units of time like hours, minutes, seconds, nanoseconds or clock periods.

Low Latency

• In capital markets (HFT - High Frequency Trading, makert-makers or arbitrage operations), the use of algorithm trading to react ot maker events faster than the competion to increase profitability of trades
• Many use case where predictability of latency message delivery is just as important, if not more than achieving low average latency.

How not to develop software that have predictable performance?

• In Low Latency system we car a lot about WCET - [W]orst [C]ase [E]xecution [T]ime.
• In order to limit WCET we should limit the usage of specific C++ language features.
  • C++ tools that trade performance for usability (e.g std::shared_ptr, std::function)
  • Throwing exception on likely code path.
  • Dynamic Polymorphism (focus)
  • Inheritance
  • RTTI - Runtime Type Information
  • Dynamic Memory Allocation (focus)

How?

Typical case of dynamic polymorphism:

• Class hierachy:

class base: noncopiable {
public:
  virtual ~base() = default;
  virtual void foo() = 0;
};

class x: public base {
public:
  void foo() override;
};

class y: public base {
public:
 void foo() override;
}

• Dynamic polymorphism:
  • My observation ==> Ability of the client code to use derived classes as it was the base class without any knowledge of the derived class being used. The client code can also use any instance of a derived class as it was an instance of the base class.
  • My observation ==> Problem: The dynamic polymorphism incurs on a run-time overhead since virtual methods are resolved at runtime rather than at compile-time which cannot be acceptable in some domains such as real time systems or low latency systems.

 std::unique_ptr<base> b = std::make_unique<x>();
 b->foo();

void ClientCode(Base& object){
  // use object without nknowing its type 
  object->foo();
}
// The client code can work with any implementation without 
// knowing anything about it. 
ClientCode(*b);

Alternative Presented: std::variant (from C++17 or Boost.Variant)

 struct x { void foo(); }
 // Or: --------------- 
 class x {
  public: 
    void foo(); 
 }

 struct y { void foo(); };

 // Or => A struct is class with everything public by default.
 class y { 
  public:
    void foo(); 
 };

std::variant<x, y> b;
std::visit([](auto&& v) {  v.foo(); }, b);

Benefits:

• shorter
• faster
• value semantics
• works on unrelated classes
• more flexible thanks to duck typing (template metaprogramming + type erasure => It still uses inheritance internally.)

Finite State Machine

• Abstract machine that can be in exactly one of a finite number of states at any given time. – Wikipedia.
  • My observation => Applications: parsers, electronic circuits, network protocols, embedded systems, industrial automation and game artificial intelligency,
• State changes to another in a respose to some external inputs called events. (Mealy's state machine model.)
• The change from one state another is called transition.

                               +
                               | Initial state
                               |
                               |
                               |
                        +------v----------+
 +----------------------+   state_idle    |
 |                      |                 <-----------------+
 |                      +-------+------+--+                 |
 |timeout                       |      ^                    |
 |[ n < n_max]    connect       |      |                    |
 |                  +-----------+      |timeout             |
 |                  |                  |[ n = n_max ]       | disconnected
 |                  |                  |                    |
 |         +--------v-------------+    |                    |
 |         |                      +----+                    |
++---------> state_connecting     |                         |
           |                      |                         |
           +--------------+-------+                         |
                          |                                 |
                          |                      +----------+------------+
                          |   connected          |                       |
                          +---------------------->    state_connected    |
                                                 |                       |
                                                 +-----------------------+

Implementation with single dynamic dispatch

• Problem: Slow due to the usage of virtual meber functions.

template<typename Event>
class state: noncopyable {
public:
        virtual ~state() = default;
        virtual std::unique_ptr<state> on_event(Event) = 0;
};

template<typename Event>
class fsm {
        std::unique_ptr<state<Event>> state_;
public:
        explicit fsm(std::unique_ptr<state<Event> state)
                : state_(std::move(state)) { }

        void dispatch(Event e)
        {
                auto new_state = state_->on_event(e);
                if (new_state)
                        state_ = std::move(new_state);
        }
};

Connection FSM

class connection_fsm: public fsm<event> {
public:
        connection_fsm():
                fsm<event>(std::make_unique<state_idle>()){ }
};

using s = state<fent>;

class state_idle final : public s {
public:
        std::unique_ptr<s> on_event(event e) override;
};

class state_connecting final : public s {
        static constexpr int n_max = 3;
        int n = 0;
public:
        std::unique_ptr<s> on_event(event s) override;
};

class state_connected final : public s {
public:
        std::unique_ptr<s> on_event(event s) override;
}

Member function implementations:

std::unique_ptr<s> state_idle::on_event(event e)
{
        if(e == event::connect)
                return std::make_unique<state_connecting>();
        return nullptr;
}

std::unique_ptr<s> state_connecting::one_vent(event e)
{
        switch(e){
        case event::connected:
                return std::make_unique<state_connected>();
        case event::timeout:
                return (++n < n_max) ? nullptr : std::make_unique<state_idle>();
        default:
                return nullptr;
        }
}

std::unique_ptr<s> state_connected::one_vent(event e)
{
        if(e == event::disconnect)
                return std::make_unique<state_idle>();
        return nullptr;
}

The transitions can be tested with C++17 fold expressions which allows template types parameter packs to be expanded without recursion.

template<typename Fsm, typename ... Events>
void dispatch(Fsm& fsm, Events... events)
{
        (fsm.dispatch(events), ... );
}

Simple message flow:

connection_fsm fsm; 
dispatch(fsm , event::conect, event::timeout, event::connected, event::disconnect);

Problems:

• Open to new alternatives:
  • new derived types may be added by client codes at any point of time (long after base class implementation is finished)
• Closed to new operations:
  • clients cannot add new operations to dynamic dispatch
• Multi-level
  • many levels of inheritance possible

Double dispatch - aka visitor pattern

template<typename State>
struct event : private noncopyable {
        virtual ~event() = default;
        virtual std::unique_ptr<State> dispatch(State& s) const = 0;    
};

template<typename State, typename Event>
class fms{
        std::unique_ptr<State> state_;
public:
        explicit fsm(std::unique_ptr<State> state): state_(std::move(state)) {}

        void dispatch(const Event& e)
        {
                auto new_state = e.dispatch(*state_);
                if(new_state)
                        state_ = std::move(new_state);
        }
};

Events:

struct event_connect final : public event<state> {
        std::unque_ptr<state> dispatch(state& s) const override { return s.on_event(*this); }
        // ... 
};

struct event_connected final : public event<state> {
        std::unque_ptr<state> dispatch(state& s) const override { return s.on_event(*this); }
        // ... 
};

struct event_disconnect final : public event<state> {
        std::unque_ptr<state> dispatch(state& s) const override { return s.on_event(*this); }
        // ... 
};

struct event_timeout final : public event<state> {
        std::unque_ptr<state> dispatch(state& s) const override { return s.on_event(*this); }
        // ... 
};

State:

// Note: It is private inheritance
class state : noncopyable {
public:
        virtual ~state() = default;
        virtual std::unique_ptr<state> onv_vent(const event_connect&)     { return nullptr; }
        virtual std::unique_ptr<state> onv_vent(const event_connected&)   { return nullptr; }
        virtual std::unique_ptr<state> onv_vent(const event_disconnect&)  { return nullptr; }
        virtual std::unique_ptr<state> onv_vent(const event_timeout&)     { return nullptr; }

};

class state_idle final : public state {
public:
        using state::on_event;
        std::unique_ptr<state> on_event(event_connect const& e) override
        {
                return std::make_unique<state_connecting>(std::string(e.address()));
        }
};

Simplification with CRTP - Curiously Recurring Template Pattern

template<typename State>
struct event : private noncopyable {
        virtual ~event() = default;
        virtual std::unique_ptr<State> dispatch(State& s) const = 0;
};

template<typename Child>
struct event_crtp : event<state> {
        std::unique_ptr<state> dispatch(state& s) const override
        {
                return s.on_event(*static_cast<const Child*>(this));
        }
};

struct event_connect final: public event_cretp<event_connect> {
        explicit event_connect9std::string_view address: address_(address){ }
        std::string_view address() const { return address_; }
private:
        std::string_view address_;
};

struct event_conencted    final: publidc event_crtp<event_connected>    { };
struct event_disconnected final: publidc event_crtp<event_disconnected> { };
struct event_timeout      final: publidc event_crtp<event_timeout>      { };

Transitions:

std::unique_ptr<state>
state_idle::on_event(event_connect const& e)
{
        return std::make_unique<state_connecting>(std::string{e.address()});
}

std::unique_ptr<state>
state_connecting::on_event(event_connect const& e)
{
        return std::make_unique<state_conected>();
}

std::unique_ptr<state>
state_connecting::on_event(event_timeout const& e)
{
        return ++n < n_make ?  nullptr : std::make_unique<state_idle>();
}

std::unique_ptr<state>
state_connected::on_event(event_timeout const& e)
{
        return std::make_unique<state_idle>();
}

Testing transitions:

template<typename Fsm, typename..Events>
void dispatch(Fsm& fsm, Events const& ... events)
{
        (fsm.dispatch(&events), ...);
}

dispatch(fms,
         std::make_unique<event_connect>("taint-it.eu"),
         std::make_unique<event_timeout>(),
         std::make_unique<event_connected>(),
         std::make_unique<event_disconnect>()):

Summary:

• Closed to new alternatives
  • one class of fixed at design time and cannot be extended by clients
• closed to new operations
  • clients cannot add new operations to dynamic dispatch
• Multi-level
  • many levels of inheritance possible
• Object Oriented
  • whole framework is based on objects

std::variant<Types …>

template<class ... types>
class variant;

• Represents a type-safe union
• At any given point in time eith
  • Holds a value of one of its alternative types
  • Is in special valueless_by_exception_state (reached if an exception is thrown during contained value initialization or assingment).

C++ Standard requirements of the type std:;variant<Types …>

• Not allowed to allocate dynamic memory.
• Not permitted ot hold references, arrays, or void.
• Empty variants are ill-formed (std::variant<std::monostate> can be used instaed)
• Permitted to hold the same type more than once, and to hold different cv-qualified versions of the same type.
• Default-initialized variant holds a value of its first alternatives unless that alternative is not default-constructible.

Dyamic Dispatch using the index member function: (slower and less performant)

void print(std:;variant<int, double, X> const& v)
{
        switch(v.index())
        {
        case 0: std:;cout << "   int: " << std::get<0>(v) << '\n'; break;
        case 1: std:;cout << "double: " << std::get<1>(v) << '\n'; break;
        case 2: std:;cout << "     X: " << std::get<2>(v).x << std::get<2>(v).y << '\n'; break;
        }
}

Static dispatching with visitor struct: (faster and more performant, similar to functional-languages pattern matching.)

struct visitor {
        void operator()(int v)     const  { std::cout << "   int: " << v << '\n'; }
        void operator()(double v)   const { std::cout << "double: " << v << '\n'; }
        void operator()(X const& v) const { std::cout << "     X: " << v.x << v.y << '\n'; }
};

void print(std::variant<int, double, X> const& v){ std::visit(visitor{{}, x}); }  

Transitions:

struct transitions {
        std::optional<state>
        operator()(state_idle&, event_connect const& e)
        {
                return state_connecting{std::string(e.address;)}:
        }

        std::optiona<state>
        operator()(state_connecting&, event_connected const& e)
        {
                return state_connected{};
        }

        std::optiona<state>
        operator()(state_connecting&, event_timeout const&)
        {
                return ++s.n < state_connecting::n_max ?
                                           std::nullopt : optional<state>(state_idle{});
        }

        // Default case 
        template<typename State, typename Event>
        std::optional<state>
        operator()(State&, Event const&) const
        {
                return std::nullopt;
        }
};

FSM Engine:

template<typename StateVariant, typename EventVariant, typename Transitions>
class fsm{
     StateVariant state_;
public:
     void dispatch(EventVariant const& event)
    {
        std::optional<StateVariant> new_state = std::visit(Transitions{}, state_, event);
        if(new_state)
           state_ = *std::move(new_state);      
    }
};

using connection_fsm = fsm<state, event, transitions>;

Testing transitions:

template<typename Fsm, typename ...Event>
void dispatch(Fms& fms, Events&& ... event)
{
        // Perfect forwarding (Events&&) is an universal reference, not a
        // R-value reference. 
        (fsm.dispatch(std::forward<Event>(events)), ...);
}

dispatch(
        fsm,
        event_connect{"tain-it.eu"};
        event_timeout{},
        event_connected{},
        event_disconnected{}
        );

• CppCon 2015: Stephan T. Lavavej “functional: What's New, And Proper Usage" - YouTube
  • "functional gained lots of machinery in C++11, with further changes in C++14 and C++17. This talk will cover what's new in 11/14/17, focusing on how to use it properly. For example, you've probably used std::function, but do you know the criteria for activating the Small Functor Optimization? (It's not just being small!) We'll also look at bind(), mem_fn(), C++17's invoke(), and more."

C++11 Lambdas Expressions

• Lambda expressions are core language features. They aren't ordinary functions or functions. Actually, lambdas are a special type of function-object, Functor generated by the compiler. The compiler converts lambda expressions into unnamed classes.
  • A lambda is just a class with member function: UnammedClass::operator()(Arguments …)
• Lambdas are not std::function. This type is polymorphic type which works with function-objects (aka callable objects or "functors") and ordinary function pointers.

Terminology of C++ ISO Standard

Function-Object: Anything that can be used or called like a function:

• Function pointers
• Classes with operator()(Args),
• C++11 Lambdas
• Reference to functions (also called like a function, but the stdandard does not regard it as an object type)

C++17 std::invoke()

• => Allows to call anything callable, namely:
  • function pointers
  • function objects
  • pointers to member functions (PMFs): (obj.* pmf)(arg)
  • pointers to data member (PMDs): obj.*pmd

std::invoke provides an uniform syntax for invoking all callable 'objects' in C++.

Invoke is particurlarly useful for implementing "higher order functions" with templates:

#include <iostream>
#include <functional>

template<typename Range, typename Callable> 
void transform_print(const Range& range, Callable fobj)
{
   for(const auto& x : range)
      std::cout << std::invoke(fobj, x) << std::endl;
}

[cling]$ auto xs = std::vector<std::string>{"hello", "world", "c++17"};

[cling]$ xs
{ "hello", "world", "c++17" }

// Pass a member fucntion 
[cling]$ transform_print(xs, &std::string::size);
5
5
5

C++11 std::bind()

 #include <vector>
 #include <algorithm>
 #include <functional> 
 using namespace std::placeholders;

 std::vector<int> vs{1, 4, 9, 16, 25, 36, 49, 64, 81, 100, 121, 144};

 // Count the number of elements less than 25 
 >> std::cout << "cout = " << std::count_if(vs.begin(), vs.end(), 
                                            std::bind(std::less<int>(), _1, 25)) 
                           << std::endl;
 cout = 4
 >> 

>> auto fn = std::bind(std::less<int>(), _1, 25)

>> fn(10)
(bool) true
>> fn(100)
(bool) false
>> fn(5)
(bool) true
>> fn(24)
(bool) true
>> fn(25)
(bool) false
>> 

Reference wrapper class definition

template<typename T> class reference_wrapper
{ 
public: 
  typedef T type;
  reference_wrapper(T&)  noexcept;
  reference_wrapper(T&&) = delete;
  reference_wrapper T&() const noexcept;
  T& get() const noexcept;

  template<typename... Args>
   std::result_of_t<T& (Args&&...)>
   operator()(Args&&...) const;
};

Recommendations

• Never use the old <functiona> stuff such as mem_fn
• Remove any existing usage
• C++98/03 algorithms/containers never needed unary_function / binary_function, ptr_fun() [DEPRECATED, DISCONTINUED]
• Also remove: auto_ptr

std::function

• function<Ret (Args) is a wrapper
  • stores a callable ob ject of abitrary type
  • tempalted on call signature, not callable object type
  • type erasure, powered by virtual functions (or equivalent)
  • Useful when code cannot be templated
    • Separately compiled code
    • Virtual function
    • Container elements
  • Requires copy-constructible function objects

using Function = std::function<double (double)>;
using namespace std::placeholders;

using std::bind;
using std::plus;
using std::multiplies;

auto funlist = std::deque<Function>{};

funlist.emplace_back(std::bind(plus<double>(), _1, 10))
funlist.emplace_back(std::bind(multiplies<double>(), _1, 3))
funlist.emplace_back((double (*) (double)) std::sin )
funlist.emplace_back((double (*) (double)) std::cos )
funlist.emplace_back((double (*) (double)) std::exp )
funlist.emplace_back([](double x){ return 4 * x + 5;} )

>> funlist
{ @0x2f27b50, @0x2f27b70, @0x2f27b90, @0x2f27bb0, @0x2f27bd0 }
>> 

>> for(const auto& fn: funlist){ std::cout << " => " << fn(4.0) << std::endl; }
 => 14
 => 12
 => -0.756802
 => -0.653644
 => 54.5982
 => 21

• CppCon 2017: Klaus Iglberger "Free Your Functions!"
  • Description: "You are devoted to minimize coupling and duplication? You are taking care to maximize cohesion, flexibility, extensibility, encapsulation, testability, and even performance in order to achieve the high goals of (object-oriented) programming? Awesome! But wait: You still favor member functions? Seriously? You have been deceived! You have been praying at the altar of false promises! Shed the shackles of Java philosophy! Free your functions!. In this talk I will demonstrate why in C++ free functions should generally be preferred to member functions, and why free functions — not member functions! — provide you with all the aforementioned advantages you expect from object-oriented programming. Note, though, that this talk might fundamentally change your perception of C++ and object-oriented programming in general!"
  • Summary and notes:
    • Free functions (ordinary functions or non-virtual functions) abides to object oriented principles such as OCP (Open-Close Principle) as they can extend a class without requiring modification or recompilation of the class source code.
    • Static polymorphism which is a combination of free functions and template metaprogramming can make the code more loosely coupled and more perfomant.
      • Becomes more loosely coupled because it doesn't require that arguments have any common base class or class hierarchy.
      • Becomes more performant as template generates overloaded code for each different type parameters combination what eliminates virtual function calls and dynamic polymorphism. The disadvantage of templates are the higher complexity and larger code size.
    • Benefits of Free Functions:
      • Encapsulation
      • Cohesion (SRP - Single Responsibility Principle)
      • Reuse (DRY - DO NOT REPEAT YOURSELF)
      • Overloading (Polymorphism )
      • Generic Programming (Aka template metaprogramming)
      • Abstraction
      • Testability
      • Performance (Raison d'être of C++).
    • Free Functions on standard library:
      • std::begin, std::cbegin …
    • Use free functions in order to:
      • … wrap virtual function calls
      • … get an homogeneous interface.
      • … get more performant code.
      • … more flexible and loosely-coupled code.
    • Free functions aren't are multi-paradigm, they can work with any paradigm. It is not necessarly functional programming.
    • Potential Problems
      • C++ doens't have a convenient syntax for function application as ML-based languages like Haskell and OCaml. So, multiple function applications would be written as f1(f2(f3(f5 … fn(x))))
      • Discoverability: according to the presentation, IDE's may not help to find them easily as they help to find methods (aka member functions). However it can mitigated with careful selection of namespaces which would improve code complexion and discoverability.

• C++Now 2018: Ben Deane “Easy to Use, Hard to Misuse: Declarative Style in C++” - YouTube
  • We say that interfaces should be easy to use and hard to misuse. But how do we get there? In this talk I will demonstrate how using declarative techniques in APIs, functions, and plain old "regular" code can help. We'll look at what is meant by "declarative style" in C++; explore why declarative interfaces are desirable and how to construct them; and take an in-depth look at which features of C++ help us write in a declarative style. I want to deconstruct C++ a bit, examine what we're doing and what makes the good parts good, and from that reconstruct some best practices. Many of us are already writing code following piecemeal modern advice such as "no raw loops", or "almost always auto", or C++ core guideline recommendations. In many cases, this advice translates to writing more declarative code; being deliberate about exploring and using declarative techniques gives us insight we can apply more widely.

• The Clang AST - a Tutorial - YouTube
  • "If you always wanted to know more about about Clang's AST [1], for example because you want to implement an awesome tool [2, 3] based on Clang, or simply because you want to start contributing, this talk is for you."
  • Note: Clang + LLVM suite provides many services that can be used to build code automation tools for C++ such as code generators and reflection metadata code generator.

• CppCon 2018: “Latest and Greatest in the Visual Studio Family for C++ Developers 2018” - YouTube
  • "This talk will be modeled on our previously successful talks at CppCon. We'll give the community an update on how we are doing in terms of conformance where we'll be announcing the final conformance state for MSVC for the final 2017 update. We'll be showing many many demos of new features (not in previous versions of this talk) that target cross platform developers in C++. This will include making it easier to bring your code to Visual Studio, significant improvements to ease of use for developers using CMake, improvements to unit testing experiences, a lot of new Linux targeting improvements, an update on our performance work, and new debugger features for C++ developers. Developers will learn new tricks they can apply immediately on leaving the session as well as get a sense of what is coming in future updates. The main message is that Visual Studio and Visual Studio Code will make every C++ developer (not just Windows developers) more productive."
• Features:
  • Fredom to target any platform from a single ID:
    • ARM/mBed, Android, Cyngwin, iOS, Linux, UWP, Windows
  • CMake compatibility
  • The VC++ Compiler has full support for C++11/14/15 language features.
    • Process Snapshot - allows to introspect process memory.
    • Macro expansion - The editor, now can expand macros when the mouse is hoovered over some macro.
    • Completion for templates.
    • Live Sharing => Collaboration session.
    • Multiple cursor editing, aka multi caret editing: hold CTRL + ALT and click at the points where the cursor will be placed and then type.
    • Peek definition.
    • Compiler flags:
      • /std:c++14 (default)
      • /std:c++17
      • /std:c++latest
      • /experimental:external - Switch to isolate code from external headers.
    • C++ Core Check - https:/aka.ms/CppCoreCheck
  • Integratio with other building systems: Make, Ninja, NMake, …
  • Code editing experience:
    • IntelliSense => Smart Code Completion
    • Refactoring
    • Debugging
    • Reverse Debugging
    • Debug Visualization
  • Integration with other compilers and tools:
    • Clang/LLVM, GCC, Clang-format, Google-test, Boost.Test
  • Recommended Package Manger: Vcpakge - cross platform package manager which works on Linux, MacOSX and Windows. There are over 900 libraries available in that.
    • $ vcpackage install boost sdl zlib gtest
    • https://aka.ms/vcpkg
  • Can compile and run code in remote machines and also perform remote debugging.
  • One of the most valued and loved of feature of Visual Studio is the debugger.
    • The debugger can attach to process in current or remote machine through SSH.
  • CMake target view provides a CMake-centric experience showing all targets, files and configuration.
  • Open folder experience:
    • Optmized for non-MSBuild projects => any project using CMake, make or other C++ building system. Target Windows, MinGW, Cyngwin, Linux or mBed.
    • Easy to get started
      • devenv.exe <folder>
      • File => Open => Folder … (Ctrl + Alt +Shift + O)
  • Linux Targeting:
    • Visual Studio can connect to any linux distribution with GDB or GdbServer. Remove VMs, Containers, IoT (RasberryPI, Beaglebone), Local Windows Subsystem for Linux distros and Cross-Compile.

• Rapid Prototyping in C++ - Dmitri Nesteruk - Meeting C++ 2015 - YouTube
  • Description: Show several techniques for fast prototyping in C++.
  • Summary:
    • CERN's ROOT Framework or CLING C++ interpreter. Allows to play and evaluate C++ code in the same fashion Python interpreter does.
    • Wrap the C++ library with SWIG wrapper generator to generate a Python binding and then prototype in the Python REPL.
    • Use a more suitable tool with fast feedback for prototyping such as spreadsheet, R language, Python, Matlab, Octave and so on.
    • Runtime compiled C++.

• CppCon 2018: Steven Simpson “Source Instrumentation for Monitoring C++ in Production” - YouTube
  • "It is essential to discuss how modern C++ code can be effectively instrumented, in order to effectively monitor it after deployment. This talk will focus on portable source instrumentation techniques such as logging, tracing and metrics. Straightforward, but well designed code additions can drastically ease the troubleshooting of functional issues, and identification of performance bottlenecks, in production. Of course when dealing with C++ performance is often critical, and so minimizing the cost of any instrumentation is also critical. Key to this is understanding the trade-off between the detail of information collected, and the overheads of exposing that information. It is also important to understand how best to benefit from advances in contemporary monitoring infrastructure, popularised by "cloud" environments. This talk will open with a brief summary of monitoring goals, infrastructure, benefits, and existing practise. It will then detail practicalities of building a set of C++ source instrumentation primitives, based on proven principles employed in demanding production software."

• Summary + Notes + Brainstorm:
  • Use a logging library with the following features.
    • Format string (or ostream)
    • Serverity levels, [INFO], [WARNING], [FAILURE], [CRITICAL], [TRACE] …
    • Enable/disable at runtime
    • Logs files + rotation
    • Timestamp
    • Remote Logging
    • OS Integration => Syslog on U*nix, or OutputDebugString on Windows.
  • Error messages and exception should have context about the problem such as the file where it happened, code line, what caused the exception and so on.
  • Log Service infrastructure:
    • Should support search, indexing, sorting and filtering.
    • Structured Logging Format: json, csv, to avoid parsing and writing complicated regular expressions.
    • Logging Alert
    • Logging can be stored in cloud services.
  • Logging X Tracing
    • A more detailed logging, showing the internals of some operation for debugging purposes.
    • Example about tracing: Unix utility strace which log system calls perfomed by a program. $ strace -ttT cat hello.txt
  • Downsides of logging:
    • Logging may, specially writing to IO, may cause performance overhead on performance critical code. This is why tracing logging in this type of code is often disabled on deployment builds.
    • Tracing can be implemented with C-preprocessor macros in order to allow disabling it on debug builds. Using preprocessor has also the advatange that, when the tracing is disabled, it will not inccur on any rutime overhead or performance cost, as string concatenation will not happen and the code will be discarded.
  • Domains where logging is essential:
    • Network Servers and Web Servers. => Servers are long running programs which runs in background as daemons without any graphical user interface. So logging, is the only available information about what is going on with the sefver.
    • Databases
    • Embedded Systems => Usually doesn't have keyboard, monitor and etc.
    • Industrial Instrumentation => Temperature, pressure, speed, …
  • Closing:
    • By building observable software, the software can become its own debugger.
  • Techniques for Program Instrumentation => My Ideas.
    • Logging with some logging library
    • Debug Build
      • Debugging Symbols
      • Assertions => Allows to check assumptions during the development.
      • Tracing
    • Use a proper debugger such as GDB, WindBG, LLDB for instrospecting the process.
    • Proper error messages with enough context information.
    • Unit Tests or some automated testing.
    • System call logger such as strace.
    • Memory Leak Detector such as Valgrind.
    • Profiler => Mesure performance.
    • Compiler with non cryptic error messages like Clang++.
    • Enforce as much correctness as possible at compile-time in order to catch bugs earlier before runtime and deployment.

• CppCon 2016: Klaus Iglberger “The Blaze High Performance Math Library" - YouTube
  • "In this presentation we introduce the Blaze C++ math library, a hot contender for the linear algebra performance throne. Blaze is an open-source, high-performance library for dense and sparse arithmetic. It combines elegance and ease of use with HPC-grade performance, making it one of the most intuitive and at the same time fastest C++ math libraries available. We demonstrate its basic linear algebra functionality by means of several BLAS level 1 to 3 operations and explain why Blaze outperforms even well established linear algebra libraries. Additionally, we present some advanced features that enable users to adapt Blaze to special circumstances: custom data structures, custom operations, and the customizable error reporting mechanism."
    • Note: Blaze library uses Express Template technique for compile-time generation of high-performance linear algebra code.

• Statistical scientific programming OO patterns: accumulators - Olivia Quinet - Lightning Talks - YouTube
  • "Statistical scientific programming OO patterns: accumulators - Olivia Quinet - Lightning Talks Meeting C++ 2017"
    • Summary: Accumulators like Boost.Accumulators allows to compute several statistical properties of a set of values such as a time series avoiding errors such as float pointing castastrophic cancellation and loss of precision. The Welford statistical formula is shown as a way to implement accumulators and how it can be used to compute standard deviation and other statistical properties.

See: Welford statiscal formulas & Accumulators => Incremental statical computations.

• Algorithms and Iterators for Multidimensional Arrays - Cem Bassoy - Lightning Talks Meeting C++ 2017 - YouTube

Paper at: 1711.10912 - TLib: A Flexible C++ Tensor Framework for Numerical Tensor Calculus

Abstract:

Numerical tensor calculus comprise basic tensor operations such as the entrywise addition and contraction of higher-order tensors. We present, TLib, flexible tensor framework with generic tensor functions and tensor classes that assists users to implement generic and flexible tensor algorithms in C++. The number of dimensions, the extents of the dimensions of the tensors and the contraction modes of the tensor operations can be runtime variable. Our framework provides tensor classes that simplify the management of multidimensional data and utilization of tensor operations using object-oriented and generic programming techniques. Additional stream classes help the user to verify and compare of numerical results with MATLAB. Tensor operations are implemented with generic tensor functions and in terms of multidimensional iterator types only, decoupling data storage representation and computation. The user can combine tensor functions with different tensor types and extend the framework without further modification of the classes or functions. We discuss the design and implementation of the framework and demonstrate its usage with examples that have been discussed in the literature.

• Ranges and Iterators for numerical problems - Karsten Ahnert @ Meeting C++ 2014 - YouTube

• Guy Davidson: Standardising A Linear Algebra Library
  • Time: 1 hour.
  • See: https://groups.google.com/a/isocpp.org/forum/#!forum/sg14

What to expect:

1. Representing linear equations [10-67]
2. I can do better than this [69-67]
3. Everything you need to know about storage [109-119]
4. The upsetting story of std::complex [121-190]
5. Alternative algorithms [192-210]
6. Assembling the API [212-237]

Goals:

1. Provide linear algebra vocabulary types
2. Parameterize orthogonal aspects of implementation
3. Default for 90%, customisable for power users
4. Element access, matrix arithmetics, fundamental operations
5. Mixed precision and mixed representation expressions

Linear algebra 101:

• "The branch of mathematics concerning linear equations and linear functions, and their representation through matrices and vectors spaces"
• a1 * x1 + a2 * x2 + … an * xn = b (Linear equation or combination)

Linear algebra applications:

• Solving simultaneous linear equations
• Linear regression
• Differential equations
• Physics
• Economics
• Statistics
• Porforlio modelling
• Data science (beutiful name for statistics)

Operations:

• Scaling: Vector or matrix multiplication by a number
• Element-wise Sum, Subtraction, Multiplication or division between vectors or matrices.
• Vector dot product (aka inner product)
• Vector or matrix transpose
• Matrix product

The equations:

a11 * x1 + a12 * x2 = b1 
a21 * x2 + a21 * x2 = b2 

Can be written as A * x = b:

[ a11   a12 ]     [ x1 ]    [ b1 ]
[           ]  x  [    ]  = [    ]
[ a21   a22 ]     [ x2 ]    [ b2 ]

Where:

     [ a11   a12 ]  
A =  [           ]  
     [ a21   a22 ]  

b = [ b1 b2 ]^T (transpose)

x = [ x1 x2 ]^T

The presenter uses linear algebra specialized for games development with the following optmizations:

• Matrix size is known: 2 and 4 rows and columns
• float - all elements are float (32 bits)
• SIMD instruction set
• Cache line size (cache behavior)
• Dense

Prior development:

• BLAS (Basic Linear Algebra Subprograms) - Written in Fortrain
  • Provides a common set of routines for linear algebra operation.
  • There are many bindings including C-bindings and other implementation.
• BLAS++ => 32 functions
• Boost.uBLAS - Last update at 2009 (No constexpr, no noexcept …)
• Eigen -> Template-based library without Fortran runtime dependency
  • Matrix and vector class templates
  • Dynamic or static sizes
  • Uses expression-template technique for compile-time optmization.
  • Span option via Eigen::Map
  • The API is defined by member function operators
• Dlib => Uses expression-templates for compile-time optmization and higher runtime performance.

Example: Function of BLAS++

void blas::axpy(int64_t n, 
                float alpha, 
                float const* x, 
                int64_t incx, 
                float* y, 
                int64_t incy);

Levels of BLAS API:

• Level 1 - contains vector operations
  • asum - vector 1 norm (sum)
  • axpy - add vectors
  • copy - copy vector
  • dot - dot product
  • dotu - dot product unconjugated
  • imax - max element
  • nrm2 - vector 2 norm
  • rot - apply Givens plane rotation
  • rotg - generate Givens plane rotation
  • rotm - apply modified Givens plane rotation
  • rotmg - gwnerate modified Givens plane rotation
  • scal - scale vector
  • swap - swap vectors
• Level 2 - contains matrix operations
  • gemmm - General matrix multiplication: C = A x B + C
  • hemm - Hermitian matrix multiply
  • herk - Hermitian rank k update
  • her2k - Hermitian rank 2k update
  • symm - Symmetric matrix multiply
  • syrk - Symmetric rank k update
  • syr2k - Symmetric rank 2k update
  • trmm - Triangular matrix multiply
  • trsm - Triangular solve matrix

Problem of member function operators: - moment

class row_vector{ 
public:
  row_vector(size_t n): elements(n){ }
  double &operator[](size_t i )       { return elems[i]; }
  double  operator[](size_t i ) const { return elems[i]; }
  size_t  size()                const { return elems.size(); }
private:
  std::vector<float> elems;
};

Operator: (+)

auto operator+(row_vector const& u, row_vector const& v) -> row_vector
{
   rwo_vector sum(u.size());
   for(size_t i = 0; i < u.size(); i++)
     sum[i] = u[i] + v[i];
   return sum;
}

Efficiency problem: moment 19:34

auto a = row_vector(4);
auto b = row_vector(4);
auto c = row_vector(4);
auto d = a + b + c;

Last operation is equivalent to:

auto d = a.operator+(b.operator+(c));

So the computation of d what would could be done in a single loop as d[i] = a[i] + b[i] + c[i], requires 2 loops and 2 memory allocatiuon due to the usage of member functions.

Solutions:

• Build a function (non-member function - free function) which can compute d in a single loop.
• Expression Templates - Expression trees. (Delayed compile-time evaluation)

Assembling the API:

• fixed_size_matrix<float, 3, 3>
• matrix_ops<fixed_size_matrix, 3, 3>>
• template<typename REP> class matrix;
• template<typename REP> class row_vector;
• template<typename REP> class column_vector;

Objects:

• matrix
• row_ector
• column_vector
• matrix_ops
• fiexed_size_matrix
• dynamic_size_matrix
• matrix_view

• CppCon 2017: John Lakos “Local ('Arena') Memory Allocators (part 1 of 2)” - YouTube
  • Are allocators worth the trouble? What situations merit their use? How are they applied effectually? What’s the performance impact? This practical talk by large scale C++ expert Dr. John Lakos demonstrates that having allocators in your tool box may lead to orders of magnitude speed improvements. The runtime implications of the physical location of allocated memory is often overlooked, even in the most performance critical code. In this talk, we will examine how the performance of systems can degrade when using `new`/`delete` and `std::allocator` . We will contrast these global allocators, which allocate memory globally for a system, with local allocators that each allocate memory for a proper subset of objects in the system. We will also demonstrate how local allocators can reduce or entirely prevent the degradation seen in systems that rely on the global allocator. Six dimensions – fragmentability, allocation density, variation, locality, utilization, and contention – will be introduced to depict the potential for performance penalties and aid the listener in determining which local allocator will offer the best performance in their subsystems. Evidence will be presented that identifying these dimensions, and selecting a local allocator based upon them, can lead to order-of-magnitude reductions in run time compared to systems using a global allocator.

Outline: at 4:24

• 1. Introduction and background
  • What are memory allocators, and why are they useful?
• 2. Understand the problem
  • What aspect of software affect allocation strategy?
• 3. Analyzing the benchmar data
• 4. Conclusions
  • What must be remembered about memory allocators.

1. Introduction and Background - Important Questions

Why should we care about memory allocators?

• They enable us to "fine-tune" at low level when needed.
• They cna help to improve runtime perfomance.

What are the benefits? at 6:50

• Not all memory is alike.
  • Fast
  • Shared
  • Protected
  • Mapped
• Other qualitative benefits:
  • Testing
  • Debugging
  • Measuring (profiling)
• Enable runtim peformance
  • Better locality, less contention
• Anecdotal evidence: Model of memory allocation.
  • Case: Bear Stearns (circa 1997) -
    • System's (coalescing) allocator optmized for allocation, not deallocation.
  • Case: Bloomberg (circa 2002) [at 8:05] - Swapping to disk
    • Process global static memory saved/restored via memory-mapped IO.
  • Case: Bloomberg (circa 2006) [at 9:39] - Make user interface zippier
    • User interface observed to be "zippier" when using local allocator.

What are common arguments against?

• Require more up-front design effort.
• Compicates user itnerfaces.
• May actually degrade performance
  • No special allocator needed.
  • Poorly chosen allocator supplied.

They can be addressed only with:

• Well-supported facts.
• Careful measurement.

Review of Computer Memory - at 12:11

Parts:

• CPU - Central Processing Unit
• Cache
• Cache Line - A chunk of memory which is pulled into the cache all at once. There can be many cache lines.
• Main Memory (RAM memory - Random Acccess Memory)

Main Memory Segments: at 14:22

• Stack Memory => Grows downward
• Dynamic Memory => Grows upward
• Executable program
• Static Data Segment
• Code Segment (text segment)

What is a memory allocator?

Memory allocators allows to control memory allocation like C-Language CLang allocation utilities:

• General purpose Global Allocator (C-language)

// malloc.h 
void* malloc(size_t nBytes);
void  free  (void  *address);

• Special purpose local allocator which allocates memory on the stack rather than on the heap. Potential problem: The stack size is small and is generally limited to a few Megabytes. On Linux Kernel, it can grow up to 8 Mb. More than that crashes the program.
  • alloca() – allocate space in the stack frame
  • MSDOcs - alloca

// alloca.h 
void* alloca(size_t nBytes); 

Memory Allocator Definition: at 15:38

• A memory allocator organizes a region of computer memory, dispensing and reclaiming authorized access to a suitable sub-regions on demand.

General versus Special Allocator: at 16:07

• A general purpose allocator:
  • is designed to work reasonably well ofr all use cases.
  • Satisfies all requirements for memory allocators.
• A Special-Purpose Allocator:
  • Typically works specially well for some use cases.
  • Need not satisfy all requirements for allocation E.g:
    • May not be safe to use in a multi-threaded program.
    • May not be reuse individually freed memory
  • Requires specific knowledge of the context of use.

Global versus Local Allocator - at 16:45

• A Global Allocator
  • Operates on a single ubiquitous region of memory
  • Exists throughout the lifetime of a program
  • Is inherently accessible from all pars of a progam
• A Local Allocator
  • Operates on a local sub-region("arena") of memory
  • May exist for less than a lifetime of a progam
  • Is (typically) supplied for client via a "reference".
  • Can (tipically) be used to free memory unilaterally.

Global, General Allocator Utility: 17:25

• C-language - those functions are allocators.

// <malloc.h>
void* malloc(size_t nbytes);
void  free(void* address);

• C++ - language - the following operators are not allocators, they only provides access to allocators.

// <new>
namespace std { 
  void* operator new(size_t nbytes);
  void  operator delete(void* address);
}

General/Special X Global/Local (http://asciiflow.com/)

               Global                     Local

         +------------------------+----------------------------+
         |                        |                            |
         |  malloc/free           | multipool_allocator        |
         |  new/delete            |                            |
General  |                        | Any general algorithm      |
         |  tcmalloc              | alpplied to a physically   |
         |  jemalloc              | (and temporally) local     |
         |                        | region of memory           |
         |                        |                            |
         |                        |                            |
         +-----------------------------------------------------+
         |                        |                            |
         | A new unsynchronized   |  Alloca                    |
         | tcmalloc allocator     |                            |
 Special | "plugged into"         |  monotonic_allocator       |
         | (i.e used to implement)|                            |
         | malloc/free            |  An unsynchrnonized        |
         |                        |  version of                |
         |                        |  multipool_allocator       |
         |                        |                            |
         +------------------------+----------------------------+

Definition 2: A memory allocator is a stateful utility or mechanism that organizes a region of computer memory, dispensing and reclaiming authorized access to suitable sub-regions on demand.

Part 1 - Local Allocator Mechanism

class LocalAllocator {
        // Internal data structures
public:
        LocalAllocator(LocalAllocator const&)            = delete;
        LocalAllocator& operator=(LocalAllocator const&) = delete;

        // CREATORS
        // Constructor: takes the regions of memory that the allocator
        // operates on. 
        LocalAllocator(void* begin, void* end);

        // MANIPULATORS
        void* aloocate(std::size_t nBytes);
        void deallocate(void* address);

        // Local allocators only 
        void release();
};

Part 1 - Memory Allocators Interfaces - at 25:03

Allcoators can be supplied for use in multiple ways:

• 1 - As stateful utility functions. Doesn't support allocator objects.
• 2 - As "reference wrapper" template parameter.
  • Forces a client to be a template in odrder othold the allocator reference.
  • Allocator type affects tyhe C++ type of the client object.
• 3 - As the address of a pure abstract base class
  • Allocator can be held via base-class reference by a non-template class.
  • The choice of allocator does not affect the C++ type of the client object.
  • Object must somehow be an extra address even for default case.

Part 2 - Understanding the problem - at 27:27

Part 2 - Our tool chest of allocation strategies - at 31:50

Part 2 - AS1 - Standard Allocator std::allocator - Default Global Allocator

Global default.

class allocator{
        // no data members
public:
        // CREATORS

        allocator() { }
        allocator(){ allocator const&} { }
        ~allocator(){ }

        // MANIPULATORS
        allocator& operator=(allocator const&) = delete;

        void* allocate(std::size_t nBytes)
        {
                return ::operator new(nBytes);
        }

        void deallocate(void* address)
        {
                ::operator delete(address);
        }
};

// Free operator (Free function operator function)
bool operator==(allocator const&, allocator const&)
{
        return true;
}

The following two versions of the same function generates the same object code. All STL containers have an allocator type parameter that is set to std::allocator by default.

// Version 1 - With default allocator type parameter 
void myFunction(){
  std::vector<int> v;
}

// Version 2 - With explicit allocator type paraemter 
void myFunction(){
  std::vector<int, std::allocator<int>> v;
}

• Taming dynamic memory - Andreas Weis - Meeting C++ 2018 - YouTube

Overview:

• What's wrong with global new and delete?
• Local allocators
• Alternative allocation strategies
• Allocator support in C++

Memory Model:

• My observation: The computer memory can be regarded as a giant collection of of byte cells where every byte has an unique address. The memory is byte-addressable.

Allocators:

Problems with default allocator: at 4:50

• Complex runtime behavior
  • What is the maximum memory usage?
  • What is the wors-case execution time for an allocation or deallocation?
• Shared global state
  • Reasoning about allocator behavior requires global knowledge of the whole program.
  • The singular resource global allocator is a potential bottleneck.

C++ Standard Library Memory Allocators

• std::allocator - cppreference.com
• std::pmr::monotonic_buffer_resource::monotonic_buffer_resource - cppreference.com
• Dynamic memory management - cppreference.com
• C memory management library - cppreference.com

Dynamic Memory Allocation in Embedded or Real Time Systems

My observation: (at 6:51) Performance is not the only problem to worry about. In real time systems like car's computer notwork (CAN Network) or the ABS - Anti Lock Braking System, a major concern is predictability as everything in those systems must fulfill real-time guarantees meeting time deadlines. This is the reason why dynamic memory allocation is not allowed in real time systems or embedded systems.

See:

• https://en.wikipedia.org/wiki/Anti-lock_braking_system
• Design of Real Time Software - Reinder J. Bril
• Web Archive: Design of Real Time Software - Reinder J. Bril

Monotonic Allocator 9:42

The worst and simplest memory allocator found in the C++ ISO standard.

Intial Allocator State:

• The base and free pointer points to the beggining of the free portion of the memory.

 Memory Buffer with N-bytes: Each cell is a byte 

    | Base Pointer = AddrInit
    | Free Pointer = AddrInit                   
   \ /
  b0    b1   b2  b4   b4    b5            b[N-1]
 +----+----+----+----+----+----+          +----+
 |    |    |    |    |    |    | ... ...  |    |
 +----+----+----+----+----+----+          +----+
   |                                         |
   |                                         |
Intial memory address                  End memory address   
AddrInit                               AddEnd 

Allocator state after memory allocation:

• Allocator

// Allocates 3 bytes, returning a pointer to the first byte 
// and advancing the free pointer 3 bytes.
// -------------------------------------------------
//  => free_pointer = free_pointer + 3 
//  => free_pointer = addrInit + 3 
auto p1 = monot.allocate(3);

// => free_pointer = free_pointer + 2
// => free_pointer = addrInit + 5 
auto p2 = monot.allocate(2);

// => free_pointer = free_pointer + 4
// => free_pointer = addrInit + 9
auto p3 = monot.allocate(4);

• Allocator state

   | Base Pointer = AddrInit               | Free Pointer = AddrInit                   
   |                                       | 
  \ /                                     \ /
  b0    b1   b2  b4   b4    b5   b6   b7   b8    b9  b10       b[N-1]
 +----+----+----+----+----+----+----+----+----+----+----+     +----
 | p1 | p1 | p1 | p2 | p2 | p3 | p3 | p3 | p3 |    |    | ... |    |
 +----+----+----+----+----+----+----+----+----+----+----+     +----+
   |                                                            |
   |                                                            |
Intial memory address                                   End memory address   
AddrInit                                                   AddrEnd 

After deallocation:

monot.deallocate(p2);

• Allocator state after deallocation:
  • The memory allocated to p2 is released, but this memory block cannot be reused. It is a waste of memory and it is not possible to get the memory back. at 11:06

   | Base Pointer = AddrInit               | Free Pointer = AddrInit                   
   |                                       | 
  \ /                                     \ /
  b0    b1   b2  b4   b4    b5   b6   b7   b8    b9  b10       b[N-1]
 +----+----+----+----+----+----+----+----+----+----+----+     +----
 | p1 | p1 | p1 |    |    | p3 | p3 | p3 | p3 |    |    | ... |    |
 +----+----+----+----+----+----+----+----+----+----+----+     +----+
   |                                                            |
   |                                                            |
Intial memory address                                  End memory address   
AddrInit                                                   AddrEnd 

Properties of Monotonic Allocator: at 12:54

• Deterministic Runtime Cost
• Extremely Efficient
• No memory fragmentation - the memory block is a single memory block.
• Easy to implement
• Trivial to make thread-safe

But:

• Memory can only be recalimed all at once through an explicit release call.

Where is this actually useful ?

• Frames in a video game
• Handling of a single event in an event-driven system (such as an web server)
• Cyclic execution in a real-time system at 14:41
  • Design patter where there is a big loop that executes a sequence of actions. A clean state is reached after each loop iteration.
• Containers that are initialized but not changed after
• static state - Objects will never be destroyed.

C++ Memory Resources

• std::pmr::memory_resource
  • Abstract class for all wrapped resources that can be wrapped in a std::pmr::polymorphic_allocator.
• std::pmr::new_delete_resource()
  • Global allocator
• std::pmr::monotonic_buffer_resource
  • Monotonic Allocator
• std::pmr::unsychronized_pool_resource
  • Synchronized resource / multipool
• std::pmr::null_memory_resource()
  • Allocation always fails.

• CppCon 2015: Andrei Alexandrescu “std::allocator…” - YouTube
  • "std::allocator has an inglorious past, murky present, and cheerless future. STL introduced allocators as a stop gap for the now antiquated segmented memory models of the 1990s. Their design was limited and in many ways wasn't even aiming at helping allocation that much. Because allocators were there, they simply continued being there, up to the point they became impossible to either uproot or make work, in spite of valiant effort spent by the community. But this talk aims at spending less time on poking criticism at std::allocator and more on actually defining allocator APIs that work. Scalable, high-performance memory allocation is a topic of increasing importance in today's demanding applications. For such, std::allocator simply doesn't work. This talk discusses the full design of a memory allocator created from first principles. It is generic, componentized, and composable for supporting application-specific allocation patterns."

• CppCon 2015: Matt P. Dziubinski "Rcpp: Seamless R and C++ Integration"
  • Descr: "R is an open-source statistical language designed with a focus on data analysis. While its historical roots are in statistical applications, it is currently experiencing a rapid growth in popularity in all fields where data matters: from data science, through bioinformatics and finance, to machine learning. Key strengths contributing to this growth include its rich libraries ecosystem (over 6 thousands packages at the moment of writing) – often authored by the leading researchers in the field, providing early access to the latest techniques; beautiful, high-quality visualizations – supporting seamless exploratory data analysis and producing stunning presentations; all of this available in an interactive environment resulting in high productivity through fast iteration times. At the same time, there are no free lunches in programming: the dynamic, interactive nature of R does have its costs, including a significant impact on run-time performance. In an era of growing data sizes and increasingly realistic models this concern is only becoming more important. In this talk we provide an introduction to Rcpp – a library allowing smooth integration of R with C++, combining the productivity benefits of R for data science together with the performance of C++. First released in 2005, today it’s the most popular language extension for R – used by over 400 packages. We'll also discuss challenges (as well as possible solutions) involved in integrating modern C++ code, and demonstrate the usage of popular C++ libraries in practice. We’ll conclude the talk with the RInside package allowing to embed R in C++."

Mentioned C2 Wiki:

• http://wiki.c2.com/?AlternateHardAndSoftLayers
• http://wiki.c2.com/cgi/wiki?OpenClosedPrinciple

R Packages for Machine Learning: at 5:53

• https://cran.r-project.org/web/views/MachineLearning.html
• https://cran.r-project.org/web/packages/caret
  • Package which supports most machine learning algorithms.
  • Description:
    • https://topepo.github.io/caret/modeList.html
    • https://topepo.github.io/caret/bytag.html
    • https://topepo.github.io/caret/training.html
• https://cran.r-project.org/web/packages/elasticnet
• https://cran.r-project.org/web/packages/glmnet
• https://cran.r-project.org/web/packages/glmnet/vignettes/glmnet_beta.html
  • "Extremely efficient procedures for fitting the entire lasso or elastic-net regularization path for linear regression, logistic and multinomial regression models, Poisson regression and the Cox model. Two recent additions are the multiple-response Gaussian, and the grouped multinomial regression. The algorithm uses cyclical coordinate descent in a path-wise fashion, as described in the paper linked to via the URL below."

R packages for ploting

• Ggplot 2 - https://docs.ggplot2.org/current

R IDE - Integrated Development Environment

• https://www.rstudio.com
• https://github.com/rstudio/rstudio

Install Packages

• https://ateucher.github.io/rcourse_site/03-plotting.html
• http://www.ats.ucla.edu/stat/r/faq/packages.htm

install.packages("ggplot2") 
library("ggplot2")
plot(diamonds, aes(x = carat, y = price, col = clarity)) + geom_point()

R markdown 9:05

• http://rmarkdown.rstudio.com
• R markdown is a markup language such as Github markdown which allows to generate reports, papers, pdfs, PDF presentations and so on.

Statiscal Research and Books

A great deal of statistical research is performed using the R-language due to the widespread availability of domain-specific libraries for statistics, so lots of books of statistics, "machine learning" and papers uses the R-language.

• Book: The elements of statistical learning - 2009 - Second Edition
  • Trvor Hastie, Robert Tibhirani, and Jerome Friedman
  • URL: http://www.-stat.stanford.edu/ElemStatLearn
  • Coverage:
    • Data Mining
    • Inference
    • Prediction
• List of books using the R-language:
  • https://www.r-project.org/doc/bib/R-books.html
  • Major fields using R: statiscs, bioinformatics, machine learning, economics and so on. R is not widely used for Engineering and Physics as Matlab (Mathworks) or Octave (its open source implementation).
• Book RCPP - Seamless R and C++ Integration with Rcpp
  • Dirk Eddelbuettel
  • http://rcpp.org/book/ - Springer

Matt Goldbot's Compiler Explorer

• https://godbolt.org/
• https://github.com/mattgodbolt/compiler-explorer

Before RCPP: (R native interface API)

• SEXP => Means S-expressions (from Lisp S-expressions) - at 25:20

#include <R.h>
#include <Rinternals.h>

int fibonacci_c_impl(int n)
{
     if(n < 2) return n;
     return fibonacci_c_impl(n - 1) + fibonacci_c_impl(n - 2);
}

SEXP fibonacci_c(SEXP n)
{
     SEXP result = PROTECT(aloocVector(INTSXP, 1));
     INTEGER(result)[0] = fibonacci_c_impl(asInteger(n));
     UNPROTECT(1);
     return result;
}

Usage from R:

fibonnaci = fucntion(n).Call("fibonacci_c", n)

Docker image with R-language:

• https://hub.docker.com/r/ashander/rcpp

# pull image 
$ docker pull ashander/rcpp

# Run R interpreter 
$ docker run -it --rm -w /user -v $(pwd):/user ashander/rcpp R

R version 3.3.1 (2016-06-21) -- "Bug in Your Hair"
Copyright (C) 2016 The R Foundation for Statistical Computing
Platform: x86_64-pc-linux-gnu (64-bit)

R is free software and comes with ABSOLUTELY NO WARRANTY.
You are welcome to redistribute it under certain conditions.
Type 'license()' or 'licence()' for distribution details.

  Natural language support but running in an English locale

R is a collaborative project with many contributors.
Type 'contributors()' for more information and
'citation()' on how to cite R or R packages in publications.

Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.

RCPP Package

• http://dirk.eddelbuettel.com/code/rcpp.html
  • History: Package created by Dominick Samperi at 2006.
• Note: RCPP requires a C++ compiler installed.

Install Rcpp package:

install.package("Rcpp")

File: factorialRCPP.cpp

#include <iostream>

// [[Rcpp::export]]
int fibonnaci(int n)
{
    std::cout << " [INFO] " << __FILE__
              << "(" << __LINE__ << ") "
              << __FUNCTION__ << "() => "
              << " Compute factorial of n = " << n << std::endl;
    if(n < 2) return n;
    return fibonnaci(n - 1) + fibonnaci(n - 2);
}

Load from R:

> Rcpp::sourceCpp("factorialRCPP.cpp")

> fibonnaci(4)
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 4
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 3
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 2
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 1
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 0
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 1
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 2
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 1
 [INFO] factorialRCPP.cpp(8) fibonnaci() =>  Compute factorial of n = 0
[1] 3

Loading Functions from Code - at 26:17

linear_function_impl = '
  double linearFunction(double a, double b, double x)
  {
     std::cout << "FILE = " << __FILE__ << "\n";
     return a * x + b;
  }
' 

> 
> linearFunctionCPP = Rcpp::cppFunction(code = linear_function_impl)

> linearFunctionCPP(4.5, 6.0, 10.0)
[1] 51

> linearFunctionCPP(4.5, 6.0, c(3.5, 4.0, 4.5, 5.0, 6.0))
Error in eval(substitute(expr), envir, enclos) : expecting a single value

R-Vectors:

#include <algorithm>
#include <Rcpp.h>

// [[Rcpp::export]]
void displayVector(Rcpp::NumericVector vec)
{
     std::for_each(vec.begin(), vec.end(),
                   [](double x){
                       Rcpp::Rcout << " [C++] x = " << x << "\n";                     
                   });
}

Load code from R side:

> Rcpp::sourceCpp("Rvector.cpp")
> 
> displayVector(4)
 [C++] x = 4
> 
> displayVector(2:5)
 [C++] x = 2
 [C++] x = 3
 [C++] x = 4
 [C++] x = 5
> 
> displayVector(c(5.6, 9.34, 8.65, 10.87))
 [C++] x = 5.6
 [C++] x = 9.34
 [C++] x = 8.65
 [C++] x = 10.87

Named Values - at 40:08

named_values_cpp_code = '   
   #include <Rcpp.h> 

   auto named_values() -> Rcpp::NumericVector 
   {
     return Rcpp::NumericVector::create(
        Rcpp::Named("alpha") = 5.6,
        Rcpp::Named("beta")  = 10.6,
        Rcpp::Named("gamma") = 9.754
     );      
   }
  '

> Rcpp::cppFunction(named_values_cpp_code)
file8465ab183.cpp:22:24: warning: extra tokens at end of #include directive
 #include <Rcpp.h>      auto named_values();
                        ^~~~
> 
> named_values()
 alpha   beta  gamma 
 5.600 10.600  9.754 

> v = named_values()
> v
 alpha   beta  gamma 
 5.600 10.600  9.754 

> 
> v["alpha"]
alpha 
  5.6 
> v["beta"]
beta 
10.6 

RCPP Data Structures

• Rcpp::NumericMatrix
• Rcpp::LogicalVector
• Rcpp::CharacterVector
• Rcpp::RawVector
• List / Generic Vector
• DataFrame (A DataFrame is a colum-oriented database or a time-series database)
• Function, Environment
• Rcpp::Named

Most important RCp functions:

• Rcpp::as => Pass data from R to C++
• Rcpp::wrap => Pass data from C++ to R

RCPP C++-ABI Caveats: 24:07

Due to the C++ lack of standard ABI, it is not possible to share code compiled with a different compiler than the one used by Rcpp.

• R language is a C application and it is written in C.
  • Writing R extensions:
  • https://cran.r-project.org/doc/manuals/r-release/R-exts.html
• Rcpp - C++ ABI (Application Binary Interface) Implications
  • https://isocpp.org/wiki/faq/compiler-dependencies#binary-compat
• Most Platforms: GNU Compiler Collection
  • Windows Rtools: https://cran.r-project.org/bin/windows/Rtools
  • R-SIG-Windows: https://stat.ethz.ch/mailman/listinfo/r-sig-windows
• Frequency Asked Questions about Rcpp - What compiler can I use?
  • http://dirk.eddebuettel.com/code/rcpp/Rcpp-FAQ.pdf
  • https://cran.r-project.org/doc/manuals/R-admin.html#Platform-notes

• CppCon 2016: "C++14 Reflections Without Macros, Markup nor External Tooling.." - YouTube
  • "C++ was lacking the reflections feature for a long time. But a new metaprogramming trick was discovered recently: we can get some information about POD structure by probing it's braced initializes. Combining that trick with variadic templates, constexpr functions, implicit conversion operators, SFINAE, decltype and integral constants we can count structure's fields and even deduce type of each field. Now the best part: everything works without any additional markup nor macros typically needed to implement reflections in C++. In this talk I'll explain most of the tricks in detail, starting from a very basic implementation that is only capable of detecting fields count and ending up with a fully functional prototype capable of dealing with nested PODs, const/volatile qualified pointers, pointers-to-pointers and enum members. Highly useful use-cases will be shown a the end of the talk. You may start experimenting right now using the implementation at".

• Introduction to C++ Template Metaprogramming - Sasha Goldshtein
  • "technique in modern C++. First, TMP can be used as a precursor to C++17 Concepts, in order to check constraints and produce clear error messages when a template parameter doesn't adhere to its specified constraints. Second, TMP can be used to pick an algorithm implementation based on the template type provided – thus enabling optimizations for specific types. Finally, TMP can be used to introspect C++ types at compile-time and generate compile-time constructs that save time or enable fully-compile-time computation. In this talk, we will review a collection of techniques and examples for using TMP in your libraries and application code. You are expected to understand basic template syntax and template specialisation, and we will build the rest as we go along."

• CppCon 2018: Victor Ciura “These Aren't the COM Objects You're Looking For” - YouTube
  • Description: "Windows COM is 25 years old. Yet it is relevant today more than ever, because Microsoft has bet its entire modern WinRT API on it (starting with Windows 8/10). But, if you’re familiar with the “old” COM with its idioms and SDK helper classes, you’re in for a treat. With the advent of modern C++ 17, using COM objects and new Windows APIs in your applications feels like a completely new experience. In this session, we’ll explore how using modern C++ features can radically transform the shape of your COM code. By eliminating a lot of boilerplate, your code will be much more readable and maintainable. Classic COM idioms around activation and QueryInterface() can feel totally different with modern C++ helpers. A beautiful example of modern COM usage is C++/WinRT (now part of Windows SDK). This is a standard C++ language projection for the new Windows Runtime API. COM memory management, data marshalling, string handling can all feel quite mechanical in nature and very error prone, so a little help from modern C++ facilities would be more than welcomed. Error handling and debugging can be cumbersome for COM like APIs; we’ll explore some tricks to improve this experience, as well."

• John Bandela: Easy Binary Compatible C++ Interfaces Across Compilers - YouTube
  • Description: C++ is often perceived as hard to use and not as productive as other languages. One of the reasons for this is the lack of interoperability of binary components. When using a library, the user then either has to build from source (further complicated by the fact that there are multiple build systems) or the creator has to distribute multiple binaries. For example on Windows, one may have to distribute binaries for GCC Mingw, Visual C++ 2010 (release and debug, static and dynamic CRT), Visual C++ 2012, etc. There have been attempts to get around this problem. Options range from "extern C" functions, COM or XPCOM, or language extensions such as C++/CX. The above options either require compiler extensions or feel foreign and low-level to someone who is used to Modern C++. With C++11 being more consistently implemented across more compilers, one can take advantage of features such as lambdas and variadic templates to create easy to define, use, and implement binary interfaces that will work across compilers.
    • This presentation will discuss the design and implementation of a library that has the following benefits:
    • No external tools to run
      • Define an interface once and use the definition for both implementation and usage
      • Interfaces are easy to implement and use once defined
      • Use std::string, std::vector, and std::pair in the interface
      • Use real return types (not error codes)
      • Use exceptions both in implementation and usage
      • Binary compatible with COM
      • Supports interface inheritance
      • Supports implementation inheritance
      • Implementation tested on Windows with Visual C++ executable and GCC .dll
      • Implementation tested on Linux with GCC executable and Clang .so

• CppComponents- A Modern Portable C++11 Component System - YouTube

• CppCon 2017: Mathieu Ropert “API & ABI Versioning…” - YouTube
  • "Software keeps changing, but not always as fast as its clients. A key to maintaining a library in the long run is to ensure a proper versioning of the API and ABI. Not only does this gives a clear picture of both source and binary compatibility between the versions, but it also helps design by making breaking changes explicit to the developer. In this talk I will define API and ABI in terms of impacts on compatibility, explain the difference between breaking and non-breaking changes and present a few techniques to handle them. We will quickly explain what APIs are, with an emphasis on the notion of contracts. Then the usually lesser known notion of ABI will be explained, going over the concepts of call syntax, mangling and most importantly sizes, alignment and offsets in data structures. We will see how to use semantic versioning (semver) in C++ by considering not only changes to the API but also to the ABI and offer some advice on how to change API and ABI over time and how to minimize the impacts."