Practical C++17

Copyright Jason Turner @le icus 1.1

 Jason Turner
Co-host of CppCast http://cppcast.com
Host of C++ Weekly https://www.youtube.com/c/JasonTurner-le icus
Co-creator of ChaiScript http://chaiscript.com
Curator of http://cppbestpractices.com
Microso MVP for C++ 2015-present

Copyright Jason Turner @le icus 1.2

 Jason Turner
Independent and available for training or contracting
http://articles.emptycrate.com/idocpp
http://articles.emptycrate.com/training.html
Next training: 2 days of best practices @ CppCon 2017

Copyright Jason Turner @le icus 1.3

 ChaiScript
Embedded scripting language co-designed by me specifically for C++
Supports Visual Studio, clang, GCC
Runs on Windows, Linux, MacOS, FreeBSD, Haiku
Currently requires C++14
Header only - no external deps
Designed for integration with C++
All types are the same and directly shared between script and C++
( double , std::string , std::function<> , etc)

Copyright Jason Turner @le icus 1.4

 ChaiScript
My proving ground for testing best practices.
About 25k lines of C++
Has evolved from C++03 + Boost -> C++11 -> C++14 and onward to
C++17
Complex template usage for automatic function type deduction

Copyright Jason Turner @le icus 1.5

 ChaiScript
Full example:
1 #include <chaiscript/chaiscript.hpp>
2
3 std::string greet(const std::string &t_name) {
4 return "Hello " + t_name;
5 }
6
7 int main() {
8 ChaiScript::chaiscript chai;
9 chai.add(chaiscript::fun(&greet), "greet");
10 chai.eval(R"(print(greet("Jason")))");
11 }

Copyright Jason Turner @le icus 1.6

 About my Talks
Move to the front!
Please interrupt and ask questions

Copyright Jason Turner @le icus 1.7

 Slack

Copyright Jason Turner @le icus 2.1

 C++17 Overview

Copyright Jason Turner @le icus 3.1

 Structured Bindings

Copyright Jason Turner @le icus 4.1

 Structured Bindings
1 auto [a,b,c] = <expression>;

Can be used to automatically split a structure into multiple variables

Added in C++17

Copyright Jason Turner @le icus 4.2

 if-init expressions

Copyright Jason Turner @le icus 5.1

 if-init expressions
1 if (auto [key, value] = *my_map.begin(); key == "mykey"){}

Added in C++17
Also works for switch conditions

Copyright Jason Turner @le icus 5.2

 Standard Library Changes

Copyright Jason Turner @le icus 6.1

 emplace_back (C++17)
1 container.emplace_back();
2 auto &val = container.back();

Becomes:
1 auto &val = container.emplace_back();

C++17 modified emplace_back and emplace_front to return references to

the objects that were just created for sequence containers.

Copyright Jason Turner @le icus 6.2

 std::string_view added in C++17
Provides a light weight std::string -like view into an array of characters.

It can be constructed from a const char * or automatically converted to

from an std::string or from a pointer and a length.
1 void use_string(std::string_view sv);

1 use_string("hello world");
2 use_string(some_std_string);

Intended to be the new language for passing string-like things.

Copyright Jason Turner @le icus 6.3

 Nested Namespaces

Copyright Jason Turner @le icus 7.1

 Nested Namespaces
C++17 adds the ability to do:
1 namespace Namespace::Nested {
2 }

Which allows for more compact namespace declarations.

Copyright Jason Turner @le icus 7.2

 Class Template Type
Deduction

Copyright Jason Turner @le icus 8.1

 Class Template Type Deduction
C++17 class template type deduction
1 template<typename First, typename Second>
2 struct Pair {
3 Pair(First t_first, Second t_second)
4 : first(std::move(t_first)), second(std::move(t_second))
5 {}
6
7 First first;
8 Second second;
9 };
10
11 int main() {
12 Pair p{1, 2.3}; /// template parameters not needed
13 }

Copyright Jason Turner @le icus 8.2

 Class Template Deduction Guides
Deduction guides tell the compiler how to deduce a class template from
a set of parameters.
1 template<typename P1, typename P2>
2 Pair(P1 &&p1, P2 &&p2) -> Pair<std::decay_t<P1>, std::decay_t<P2>>;

Copyright Jason Turner @le icus 8.3

 Class Template Deduction Guides
Which helps for cases like this:
1 template<typename First, typename Second>
2 struct Pair {
3 template<typename P1, typename P2>
4 Pair(P1 &&p1, P2 &&p2)
5 : first(std::forward<P1>(p1)), second(std::forward<P2>(p2))
6 {}
7
8 First first;
9 Second second;
10 };
11
12 template<typename P1, typename P2> ///
13 Pair(P1 &&p1, P2 &&p2) -> Pair<std::decay_t<P1>, std::decay_t<P2>>; ///
14
15 int main() {
16 Pair p{1, 2.3}; /// deduction now succeeds
17 }

Copyright Jason Turner @le icus 8.4

 if constexpr

Copyright Jason Turner @le icus 9.1

 if constexpr
A compile-time conditional block
1 if constexpr( /* constant expression */ ) {
2 // if true, this block is compiled
3 } else {
4 // if false, this block is compiled
5 }

Copyright Jason Turner @le icus 9.2

 Fold Expressions

Copyright Jason Turner @le icus 10.1

 Fold Expressions
1 ( ... <op> <pack expression> ) // unary left fold
2 ( <pack expression> <op> ... ) // unary right fold
3 ( <init> <op> ... <op> <pack expression> ) // binary left fold
4 ( <pack expression> <op> ... <op> <init> ) // binary right fold

Allows for "folding" of a variadic parameter pack into a single value.

Allowed operations:
+ - * / % ^ & | << >> += -= *= /= %= ^= &= |= <<= >>= = == != < >
<= >= && || , .* ->*

Copyright Jason Turner @le icus 10.2

 Unary le fold
1 ( ... <op> <pack expression> ) // unary left fold

1 ( ... && args )

Expands to:
1 ( ( arg1 && arg2 ) && arg3 )

Copyright Jason Turner @le icus 10.3

 Unary right fold
1 ( <pack expression> <op> ... ) // unary right fold

1 ( args && ... )

Expands to:
1 ( arg1 && ( arg2 && arg3 ) )

Copyright Jason Turner @le icus 10.4

 Binary le fold
1 ( <init> <op> ... <op> <pack expression> )

1 ( true && ... && args )

Expands to:
1 ( ( ( true && arg1 ) && arg2 ) && arg3 )

Copyright Jason Turner @le icus 10.5

 Binary right fold
1 ( <pack expression> <op> ... <op> <init> )

1 ( args && ... && true )

Expands to:
1 ( arg1 && ( arg2 && ( arg3 && true ) ) )

Copyright Jason Turner @le icus 10.6

 noexcept is now part of the
type system

Copyright Jason Turner @le icus 11.1

 noexcept In The Type System
1 void use_func(void (*func)() noexcept);
2 void my_func();
3
4 use_func(&my_func);

This will no longer compile in C++17, a new overload is needed:

1 void use_func(void (*func)());

noexceptfunction pointer can be converted to non- noexcept , but not the

other way around.

Copyright Jason Turner @le icus 11.2

 Feature Impact

Copyright Jason Turner @le icus 12.1

 Which C++ 17 Had The Most Impact on
ChaiScript?
What is important for defining impact?
Compile Time?
Runtime?
Code Size?
Readability?
I tried to review each of these.

Copyright Jason Turner @le icus 12.2

 Which C++ 17 Had The Most Impact on
ChaiScript?
Vote https://strawpoll.com/w767k3kk:
Class template type deduction / guides?
if / switch init expressions?
fold expressions?
if constexpr ?
structured bindings?
std::string_view
std::vector::emplace_back
nested namespace s
noexcept in the type system
Copyright Jason Turner @le icus 12.3
 Results Least to Greatest

Copyright Jason Turner @le icus 13.1

noexcept In The Type System

Copyright Jason Turner @le icus 14.1

 Impact of noexcept in the type system
Caused a 2x duplication of my function type deduction code
Required conditional compilation with
#if __cpp_noexcept_function_type >= 201510 to get it to compile in
C++17 at all.
This typed support for noexcept appears to give ~1% performance
boost

Copyright Jason Turner @le icus 14.2

 Impact of noexcept in the type system
Impact conclusions:
Performance: Possible Minor Improvement
Maintainability: Negative: more overloads necessary, conditional code
Compile Time: None observed
Readability: None
Score: -0.5 necessary but potentially annoying change.

Copyright Jason Turner @le icus 14.3

 if-init Expressions

Copyright Jason Turner @le icus 15.1

 if-init Expression Impact
I was excited about the possibility of limiting variable scope:
1 auto val = get_some_val();
2 if (val > 5) {
3 // do something else
4 }

becomes
1 if (auto val = get_some_val(); val > 5) {
2 // do something else
3 }

Copyright Jason Turner @le icus 15.2

 if-init Expression Impact
But if your code is well modularized and has small functions...
1 bool my_func() {
2 auto val = get_some_val();
3 if (val > 5) {
4 return something(val);
5 } else {
6 return something_else(val);
7 }
8 }

vs
1 bool my_func() {
2 if (auto val = get_some_val(); val > 5) {
3 return something(val);
4 } else {
5 return something_else(val);
6 }
7 }
Copyright Jason Turner @le icus 15.3
 if-init Expression Impact
Then how do you format it for readability?
1 bool my_func() {
2 if (auto val = get_some_val(); val > 5) {
3 return something(val);
4 } else {
5 return something_else(val);
6 }
7 }

vs
1 bool my_func() {
2 if (auto val = get_some_val();
3 val > 5) {
4 return something(val);
5 } else {
6 return something_else(val);
7 }
8 }
Copyright Jason Turner @le icus 15.4
 if-init Expression Impact
Impact conclusions:
Performance: None
Maintainability: Small
Compile Time: None
Readability: Possibly Hurt
Score: 0.5 disappointing

Copyright Jason Turner @le icus 15.5

 Nested Namespaces

Copyright Jason Turner @le icus 16.1

 Impact of nested namespaces
1 namespace chaiscript {
2 namespace dispatch {
3 namespace detail {
4 // some thing
5 }
6 // some other thing
7 }
8 }

becomes
1 namespace chaiscript::dispatch {
2 namespace detail {
3 // some thing
4 }
5 // some other thing
6 }

Copyright Jason Turner @le icus 16.2

 Impact of nested namespaces
I found these changes quite satisfying to implement.
However:
Causes large whitespace diffs
Can cause large diffs again if you need to insert something in a middle
namespace

Copyright Jason Turner @le icus 16.3

 Impact of nested namespaces
Impact conclusions:
Performance: none
Maintainability: possibly negative
Compile time: none
Readability: large
Score: 1.5 surprisingly gratifying to implement. Might make us rethink
some of our coding and formatting standards

Copyright Jason Turner @le icus 16.4

 Class Template Type
Deduction

Copyright Jason Turner @le icus 17.1

 Class Template Type Deduction Impact
C++17 adds a complete complement of deduction guides for standard
library types.
What impact did this have on my existing code base?
Almost 0

Copyright Jason Turner @le icus 17.2

 Class Template Type Deduction Impact
Class template type deduction doesn't help at the class level
1 struct S
2 {
3 // no practical way to use type deduction here
4 std::vector<int> myvec;
5 };

Copyright Jason Turner @le icus 17.3

 Class Template Type Deduction Impact
I almost never create local containers, and auto picks up the rest.
Which is better (pretend they do more)?
1 void myfunc() {
2 /// A
3 auto f = [](){ /* something */ };
4 f();
5 }

1 void myfunc() {
2 /// B
3 std::function f = [](){ /* something */ };
4 f();
5 }

Copyright Jason Turner @le icus 17.4

 Class Template Type Deduction Impact
Who thinks these are good uses of class template type deduction?
1 std::vector some_strings{
2 "This is a list"s,
3 "of somewhat longish"s,
4 "strings to operate on."s
5 };

1 std::vector some_objs{
2 std::make_unique<Obj>(1),
3 std::make_unique<Obj>(2),
4 std::make_unique<Obj>(3),
5 };

Copyright Jason Turner @le icus 17.5

 Class Template Type Deduction Impact
Impact conclusions:
Performance: Possible negative if not considering object lifetime
Maintainability: Small
Compile Time: None
Readability: Small
Score: 1.5

Copyright Jason Turner @le icus 17.6

 Structured Bindings

Copyright Jason Turner @le icus 18.1

 Impact of structured bindings
Good, helps clean up code.
1 void dump_types() const {
2 for (const auto &v : get_types()) {
3 std::cout << v.first << ": " << v.second->name() << '\n';
4 }
5 }

becomes
1 void dump_types() const {
2 for (const auto &[name, type] : get_types()) {
3 std::cout << name << ": " << type->name() << '\n';
4 }
5 }

Copyright Jason Turner @le icus 18.2

 Impact of structured bindings
Is there any performance impact?

Copyright Jason Turner @le icus 18.3

 Impact of structured bindings
Which is better?
1 std::string func() {
2 auto val = get_some_val();
3 return val;
4 }

or
1 std::string func() {
2 auto val = get_some_val();
3 return std::move(val);
4 }

Copyright Jason Turner @le icus 18.4

 Impact of structured bindings
Why?
1 std::string func() {
2 auto val = get_some_val();
3 return val;
4 }

or
1 std::string func() {
2 auto val = get_some_val();
3 return std::move(val); /// move forced, breaks NRVO
4 }

Copyright Jason Turner @le icus 18.5

 Impact of structured bindings
Which is better?
1 std::string func() {
2 auto pair = get_some_pair();
3 return pair.second;
4 }

or
1 std::string func() {
2 auto pair = get_some_pair();
3 return std::move(pair.second);
4 }

Copyright Jason Turner @le icus 18.6

 Impact of structured bindings
Why?
1 std::string func() {
2 auto pair = get_some_pair();
3 return pair.second; /// NRVO cannot apply to subobject
4 }

or
1 std::string func() {
2 auto pair = get_some_pair();
3 return std::move(pair.second);
4 }

Copyright Jason Turner @le icus 18.7

 Impact of structured bindings
Which is better?
1 std::string func() {
2 auto [succeeded, value] = get_some_pair();
3 return value;
4 }

or
1 std::string func() {
2 auto [succeeded, value] = get_some_pair();
3 return std::move(value);
4 }

Copyright Jason Turner @le icus 18.8

 Impact of structured bindings
Why?
1 std::string func() {
2 auto [succeeded, value] = get_some_pair();
3 return value; /// NRVO cannot apply to sub object
4 }

or
1 std::string func() {
2 auto [succeeded, value] = get_some_pair();
3 return std::move(value);
4 }

Copyright Jason Turner @le icus 18.9

 Impact of structured bindings
This is (effectively) what the compiler is doing for us
1 std::string func() {
2 auto e = get_some_pair();
3 auto &succeeded = e.first;
4 auto &value = e.second;
5 return value;
6 }

Copyright Jason Turner @le icus 18.10

 Impact of structured bindings
Impact conclusions:
Performance: potentially very negative. Now you have to consider the
source of the variable.
Maintainability: Minor
Compile Time: None
Readability: Large
Score: 2 potentially really helpful, but mess with our understand of
object lifetime

Copyright Jason Turner @le icus 18.11

 Fold Expressions

Copyright Jason Turner @le icus 19.1

 Fold Expression Impact
Almost everything that is possible with fold expressions, is possible with
initializer_lists
1 (void)std::initializer_list<int>{
2 (static_cast<T&>(*this).trace(ds, node), 0)...
3 };

became
1 (static_cast<T&>(*this).trace(ds, node), ... );

Copyright Jason Turner @le icus 19.2

 Fold Expression Impact
Everyone agrees both versions have a well defined order, right?
1 (void)std::initializer_list<int>{
2 (static_cast<T&>(*this).trace(ds, node), 0)...
3 };

became
1 (static_cast<T&>(*this).trace(ds, node), ... );

Copyright Jason Turner @le icus 19.3

 Fold Expression Impact
Used in this context:
1 void do_trace(const Dispatch_State &ds,
2 const AST_Node_Impl<Tracer<T...>> *node)
3 {
4 (static_cast<T&>(*this).trace(ds, node), ... );
5 }

Copyright Jason Turner @le icus 19.4

 Fold Expression Impact
Impact conclusions:
Performance: None
Maintainability: Small
Compile Time: None
Readability: Small
Score 2: no negatives, minor help in some cases

Copyright Jason Turner @le icus 19.5

 Emplace Back

Copyright Jason Turner @le icus 20.1

 Impact of C++17 emplace_back
This:
1 Boxed_Value &add_get_object(const std::string &t_name,
2 Boxed_Value obj, Stack_Holder &t_holder)
3 {
4 auto &stack_elem = get_stack_data(t_holder).back();
5
6 if (std::any_of(stack_elem.begin(), stack_elem.end(),
7 [&](const std::pair<std::string, Boxed_Value> &o) {
8 return o.first == t_name;
9 }))
10 {
11 throw chaiscript::exception::name_conflict_error(t_name);
12 }
13
14 stack_elem.emplace_back(t_name, std::move(obj))
15 return stack_elem.back().second;
16 }

Copyright Jason Turner @le icus 20.2

 Impact of C++17 emplace_back
Became:
1 Boxed_Value &add_get_object(const std::string &t_name,
2 Boxed_Value obj, Stack_Holder &t_holder)
3 {
4 auto &stack_elem = get_stack_data(t_holder).back();
5
6 if (std::any_of(stack_elem.begin(), stack_elem.end(),
7 [&](const std::pair<std::string, Boxed_Value> &o) {
8 return o.first == t_name;
9 }))
10 {
11 throw chaiscript::exception::name_conflict_error(t_name);
12 }
13
14 return stack_elem.emplace_back(t_name, std::move(obj)).second;
15 }

Copyright Jason Turner @le icus 20.3

 Impact of C++17 emplace_back
Impact Conclusions:
Performance: tiny possibility of performance improvement (I did
measure some)
Maintainability: minor
Compile time: none
Readability: minor
Score: 2.5 possible performance, readability, and maintainability helps

Copyright Jason Turner @le icus 20.4

 std::string_view

Copyright Jason Turner @le icus 21.1

 Impact of std::string_view
The main impact of using std::string_view is in looking up of identifiers
and avoiding temporary strings.
But this comes with a complexity cost.

Copyright Jason Turner @le icus 21.2

 Impact of std::string_view
Cost of std::string_view :

What does this string_view constructor need to do?

1 std::string_view(const char *);

It must calculate the length of the passed in string. Which might happen
at compile-time or might not.

Copyright Jason Turner @le icus 21.3

 Impact of std::string_view
Provided conversions:
std::string automatically converts to std::string_view via conversion
operator
const char * automatically converts to std::string via non- explicit
constructor
const char * automatically converts to std::string_view via
non- explicit constructor
Provided comparisons:
operator<(const std::string &lhs, const std::string &rhs)
operator<(std::string_view lhs, std::string_view rhs)
etc
Copyright Jason Turner @le icus 21.4
 Impact of std::string_view
Is this code OK?
1 std::map<std::string_view, int> map;
2
3 void add(const std::string &t_str, int val)
4 {
5 map[t_str] = val;
6 }

Copyright Jason Turner @le icus 21.5

 Impact of std::string_view
Bad - possible pointer to temporary stored
1 std::map<std::string_view, int> map;
2
3 void add(const std::string &t_str, int val)
4 {
5 map[t_str] = val; ///
6 }

Copyright Jason Turner @le icus 21.6

 Impact of std::string_view
Storing a map of string_view is very risky, and you won't get any compiler
warnings.

Copyright Jason Turner @le icus 21.7

 Impact of std::string_view
Does this compile?
1 std::map<std::string, int> map;
2
3 int get(std::string_view t_sv) {
4 return map.at(t_sv);
5 }

Copyright Jason Turner @le icus 21.8

 Impact of std::string_view
Does this compile? No.
1 std::map<std::string, int> map;
2
3 int get(std::string_view t_sv) {
4 return map.at(t_sv); // no conversion to std::string
5 }

Copyright Jason Turner @le icus 21.9

 Impact of std::string_view
We need to use is_transparent comparitor

Copyright Jason Turner @le icus 21.10

 Impact of std::string_view
is_transparent is provided by std::less<>

1 std::map<std::string, int, std::less<>> map;

2
3 int get(std::string_view t_sv) {
4
5
6 if (auto itr = map.find(t_sv); itr != map.end()) {
7 return *itr;
8 }
9 throw std::range_error("Key not found");
10 }

Copyright Jason Turner @le icus 21.11

 Impact of std::string_view
is_transparent is provided by std::less<>

1 std::map<std::string, int, std::less<>> map;

2
3 int get(std::string_view t_sv) {
4
5 /// What happens for this find?
6 if (auto itr = map.find(t_sv); itr != map.end()) {
7 return *itr;
8 }
9 throw std::range_error("Key not found");
10 }

Copyright Jason Turner @le icus 21.12

 Impact of std::string_view
is_transparent is provided by std::less<>

1 std::map<std::string, int, std::less<>> map;

2
3 int get(std::string_view t_sv) {
4 /// a temporary string_view is need for each comparison
5 /// by using the operator<(string_view, string_view)
6 if (auto itr = map.find(t_sv); itr != map.end()) {
7 return *itr;
8 }
9 throw std::range_error("Key not found");
10 }

Copyright Jason Turner @le icus 21.13

 Impact of std::string_view
is_transparent is provided by std::less<>

1 std::map<std::string, int, std::less<>> map;

2
3 int get(std::string_view t_sv) {
4
5 /// How do we do better?
6 if (auto itr = map.find(t_sv); itr != map.end()) {
7 return *itr;
8 }
9 throw std::range_error("Key not found");
10 }

Copyright Jason Turner @le icus 21.14

 Impact of std::string_view
std::less<void> looks something like this:
1 struct less {
2 template<typename LHS, typename RHS>
3 constexpr bool operator()(const LHS &lhs, const RHS &rhs) const
4 {
5 return lhs < rhs;
6 }
7 struct is_tansparent{};
8 };

Copyright Jason Turner @le icus 21.15

 Impact of std::string_view
My version is:
1 struct str_less {
2 constexpr bool operator()(
3 const std::string &t_lhs, const std::string &t_rhs) const noexcept
4 {
5 return t_lhs < t_rhs;
6 }
7
8 template<typename LHS, typename RHS>
9 constexpr bool operator()(
10 const LHS &t_lhs, const RHS &t_rhs) const noexcept
11 {
12 return std::lexicographical_compare(
13 t_lhs.begin(), t_lhs.end(), t_rhs.begin(), t_rhs.end());
14 }
15 struct is_transparent{};
16 };

Copyright Jason Turner @le icus 21.16

 Impact of std::string_view
I'm sure this is highly dependent on compiler and optimization settings.
This was G++7.2 for me.
1 /// using my str_less saved 2% performance
2 /// across ChaiScript
3 std::map<std::string, int, str_less> map;
4
5 int get(std::string_view t_sv) {
6 if (auto itr = map.find(t_sv); itr != map.end()) {
7 return *itr;
8 }
9 throw std::range_error("Key not found");
10 }

Copyright Jason Turner @le icus 21.17

 Impact of std::string_view
Also, don't do this:
1 struct S {
2 S(std::string_view sv) : m_string(std::move(sv)) {}
3 std::string m_string;
4 };

Copyright Jason Turner @le icus 21.18

 Impact of std::string_view
Or worse:
1 S make_s(std::string s) {
2 return S(std::move(s));
3 }
4
5 struct S {
6 S(std::string_view sv) : m_string(std::move(sv)) {}
7 std::string m_string;
8 };

Copyright Jason Turner @le icus 21.19

 Impact of std::string_view
Which can become: (not actually contrived)
1 S make_s(std::string s) {
2 return S(std::move(s));
3 }
4
5 struct S {
6 S(std::string_view sv) : m_string(std::move(sv)) {}
7 std::string m_string;
8 };
9
10 make_s("Hello World");

Copyright Jason Turner @le icus 21.20

 Impact of std::string_view
Performance: Possibly large either good or bad
Maintainability: minor: remove_prefix is helpful with parsing
Compile time: minor
Readability: minor (you know that it's only a "view")
Score: 3 can help in virtually every way, but be aware of pitfalls

Copyright Jason Turner @le icus 21.21

 Impact of std::string_view
Notes for deployment:
Use consistently when you move to it, otherwise expensive
conversions back and forth
If you know you're going to ultimately need a std::string use
std::string through the whole call chain
Can have huge impact if you have string data that can be constexpr
I'm still seeing buggy / incomplete constexpr implementations
Understand the impact of using string_view with lookup for std::map /
std::set keys

Copyright Jason Turner @le icus 21.22

 if constexpr

Copyright Jason Turner @le icus 22.1

 if constexpr Impact
This:
1 template<typename Type>
2 auto do_call_impl(Class *o) const
3 -> std::enable_if_t<std::is_pointer_v<Type>, Boxed_Value>
4 {
5 return Handle_Return<Type>
6 ::handle(o->*m_attr);
7 }
8
9 template<typename Type>
10 auto do_call_impl(Class *o) const
11 -> std::enable_if_t<!std::is_pointer_v<Type>, Boxed_Value>
12 {
13 return Handle_Return<std::add_lvalue_reference_t<Type>>
14 ::handle(o->*m_attr);
15 }

Copyright Jason Turner @le icus 22.2

 if constexpr Impact
Can become:
1 template<typename Type>
2 auto do_call_impl(Class *o) const
3 {
4 if constexpr(std::is_pointer_v<Type>) {
5 return detail::Handle_Return<Type>
6 ::handle(o->*m_attr);
7 } else {
8 return detail::Handle_Return<std::add_lvalue_reference_t<Type>>
9 ::handle(o->*m_attr);
10 }
11 }

Copyright Jason Turner @le icus 22.3

 if constexpr Impact
Which eliminated about 1/2 of my SFINAE usage, but why not 100%?
constructors and other mixed templated / typed overloads

Copyright Jason Turner @le icus 22.4

 if constexpr Impact
1 Any(const Any &t_any)
2 : m_data(t_any.empty() ? nullptr : t_any.m_data->clone())
3 { }
4
5 template<typename ValueType,
6 typename = std::enable_if_t<
7 !std::is_same_v<Any, std::decay_t<ValueType>>>>
8 explicit Any(ValueType &&t_value)
9 : m_data(std::make_unique<Data_Impl<std::decay_t<ValueType>>>(
10 std::forward<ValueType>(t_value)))
11 {
12 }

Copyright Jason Turner @le icus 22.5

 if constexpr Impact
1 template<Typename T>
2 static std::unique_ptr<Data> make_data(T &&t) {
3 if constexpr (std::is_same_v<Any, std::decay_t<T>>) {
4 if (!t_any.empty()) { return t_any.m_data->clone(); }
5 else { return nullptr; }
6 } else {
7 return make_unique<Data_Impl<std::decay_t<ValueType>>>(
8 std::forward<ValueType>(t_value));
9 }
10 }
11
12 template<typename ValueType> explicit Any(ValueType &&t_value)
13 : m_data(make_data(std::forward<ValueType>(t_value)))
14 { }

What problem (if any) does this version have?

Copyright Jason Turner @le icus 22.6

 if constexpr Impact
Covering move construction
1 template<Typename T> std::unique_ptr<Data> make_data(T &&t) {
2 if constexpr (std::is_same_v<Any, std::decay_t<T>>) {
3 if constexpr (std::!is_const_v<decltype(t)>
4 && std::is_rvalue_reference_v<decltype(t)>) {
5 return std::move(t.m_data);
6 } else {
7 if (!t_any.empty()) { return t_any.m_data->clone(); }
8 else { return nullptr; }
9 }
10 } else {
11 return make_unique<Data_Impl<std::decay_t<ValueType>>>
12 (std::forward<ValueType>(t_value))
13 }
14 }
15
16 template<typename ValueType> explicit Any(ValueType &&t_value)
17 : m_data(make_data(std::forward<ValueType>(t_value)))
18 { }

Copyright Jason Turner @le icus 22.7

 if constexpr Impact
Now, I don't like SFINAE, but I'm sticking with it for eliminating options
for overload resolution.

Copyright Jason Turner @le icus 22.8

 if constexpr Impact
Side note: Boxed_Number
ChaiScript allows for all operators for all possible arithmetic types
While maintaining the proper types, following C++ standards
Example: float * int returns float , uint64_t * int returns uint64_t
This means all possible combinations must be generated at compile
time
if constexprenabled a rewrite of this code which removed about 160
LOC and allowed for more linear flow

Copyright Jason Turner @le icus 22.9

 if constexpr Impact
Impact conclusions:
Performance: none
Maintainability: large
Compile Time: minor (for my use cases)
Readability: large
Score: 5 generally very helpful, no real caveats.

Copyright Jason Turner @le icus 22.10

 Class Template Deduction
Guides

Copyright Jason Turner @le icus 23.1

 Class Template Deduction Guide Impact
While class template type deduction specifically had almost no impact,
deduction guides had a significant impact

Copyright Jason Turner @le icus 23.2

 Class Template Deduction Guide Impact
This:
1 template<typename Ret, typename ... Param>
2 Function fun(Ret (*func)(Param...)) {
3 auto fun_call = Fun_Caller<Ret, Param...>(func);
4 return Function(make_shared<Function_Callable_Impl<
5 Ret (Param...), decltype(fun_call)>>(fun_call));
6 }
7
8 template<typename Ret, typename Class, typename ... Param>
9 Function fun(Ret (Class::*t_func)(Param...) const) {
10 auto call = Const_Caller<Ret, Class, Param...>(t_func);
11 return Function(make_shared<Function_Callable_Impl<
12 Ret (const Class &, Param...), decltype(call)>>(call));
13 }

Plus noexcept , non- const etc

Copyright Jason Turner @le icus 23.3

 Class Template Deduction Guide Impact
Is replaced with a series of deduction guides:
1 template<typename Ret, typename Class, typename ... Param>
2 Signature(Ret (Class::*f)(Param ...) volatile) ->
3 Signature<Ret, Params<volatile Class &, Param...>, false, true>;
4
5 template<typename Ret, typename Class, typename ... Param>
6 Signature(Ret (Class::*f)(Param ...) volatile noexcept) ->
7 Signature<Ret, Params<volatile Class &, Param...>, true, true>;
8
9 template<typename Ret, typename Class, typename ... Param>
10 Signature(Ret (Class::*f)(Param ...) volatile const) ->
11 Signature<Ret, Params<volatile const Class &, Param...>, false, true>;
12
13 template<typename Ret, typename Class, typename ... Param>
14 Signature(Ret (Class::*f)(Param ...) volatile const noexcept) ->
15 Signature<Ret, Params<volatile const Class &, Param...>, true, true>;

Plus one simple function

Copyright Jason Turner @le icus 23.4

 Class Template Deduction Guide Impact
Net result:
no change in LOC
8x better support for C++ functions ( && , & , volatile )
eliminated mostly redundant code.

Copyright Jason Turner @le icus 23.5

 Class Template Deduction Guide Impact
Impact conclusions:
Performance: Minor
Maintainability: Large
Compile Time: Minor
Readability: Large
Score: 6 possibility of large help, including performance increases with
rewritten code

Copyright Jason Turner @le icus 23.6

Copyright Jason Turner Conclusions
@le icus

Copyright Jason Turner @le icus 24.1

 Conclusions
There are many constexpr changes which we did not address
The little details have the biggest impact
[[maybe_unused]] gets an honorable mention for removing code like
(void)variable
Being able to do and
std::decay_t<Type>
std::is_trivially_destructible_v<Type> instead of
typename std::decay<Type>::type and
std::is_trivially_destructible<Type>::value goes a long way to
making templated code more readable

Copyright Jason Turner @le icus 24.2

 Jason Turner
Co-host of CppCast http://cppcast.com
Host of C++ Weekly https://www.youtube.com/c/JasonTurner-le icus
Co-creator of ChaiScript http://chaiscript.com
Curator of http://cppbestpractices.com
Microso MVP for C++ 2015-present

Copyright Jason Turner @le icus 24.3

 Jason Turner
Independent and available for training or contracting
http://articles.emptycrate.com/idocpp
http://articles.emptycrate.com/training.html
Next training: 2 days of best practices @ CppCon 2017

Copyright Jason Turner @le icus 24.4