スライドショー:: イテレータからレンジまで

class: middle

# From Iterators To Ranges
## The Upcoming Evolution Of the Standard Library

---

# Questions

- Who knows Boost.Range?
- Who knows Eric Niebler's Range V3 library?
- Who uses ranges in everyday programming?

---

# Why Ranges?

```cpp
std::vector<T> vec=...;
std::sort( vec.begin(), vec.end() );
vec.erase( std::unique( vec.begin, vec.end() ), vec.end() );
```

How often do we have to mention `vec`?

Pairs of iterators belong together -> put into one object!

```cpp
tc::unique_inplace(tc::sort(vec));
```

Much nicer!

---

# Why do I think I know something about ranges?

- think-cell has a range library
  * evolved from Boost.Range
- 1 million lines of production code use it
- Library and production code evolve together
  * ready to change library and production code anytime
  * no obstacle to library design changes
  * large code base to try them out

---

# Ranges in C++11

- range-based for
```cpp
for( int& i : <range_expression> ) {
	...	
}
```

- universal access to begin/end
```cpp
std::begin/end(<range_expression>)
```

- that's all ... :-(

---

#Future of Ranges

- Eric Niebler's pet project

- Ranges TS ("Technical Specification")
  * `<algorithm>` supports ranges

```cpp
namespace ranges {
	template< typename Rng, typename What >
	decltype(auto) find( Rng && rng, What const& what ) {
		return std::find(
			ranges::begin(rng),
			ranges::end(rng),
			what
		);
	}
}
```

- Range V3
  * preview of what Eric wants to standardize

---

#What are Ranges?

- Containers
```
vector
string
list
```
   * own elements
   * deep copying
          * copying copies elements in O(N)
   * deep constness
          * `const` objects implies `const` elements

--
  
- Views
  
  
```
								 Range
                                  /\				
                                 /  \
                                /    \
                          Container  View
```
  
---

#Views

```cpp
template<typename It>
struct iterator_range {
	It m_itBegin;
	It m_itEnd;
	It begin() const {
		return m_itBegin;
	}
	It end() const {
		return m_itEnd;
	}
};
```

- reference elements
- shallow copying
  * copying copies reference in O(1)
- shallow constness
  * view object `const` independent of element `const`

---

#More Interesting Views: Range Adaptors

```cpp
std::vector<int> v;
auto it=ranges::find(
	v,
	4
); // first element of value 4.
```
vs.
```cpp
struct A {
	int id;
	double data;
};
std::vector<A> v={...};
auto it=ranges::find_if(
	v,
	[](A const& a){ return a.id==4; } // first element of value 4 in id
);
```

- Similar in semantics
- Not at all similar in syntax

---

#Transform Adaptor

```cpp
std::vector<int> v;
auto it=ranges::find(
	v,
	4
); // first element of value 4.
```
vs.
```cpp
struct A {
	int id;
	double data;
};
std::vector<A> v={...};
auto it=ranges::find(
	tc::transform(v,std::mem_fn(&A::id)),
	4
); // first element of value 4 in id
```

---

#Transform Adaptor (2)

```cpp
struct A {
	int id;
	double data;
};
std::vector<A> v={...};
auto it=ranges::find(
	tc::transform(v,std::mem_fn(&A::id)),
	4
); // first element of value 4 in id
```

What is `it` pointing to?

- `int`!

--
	
What if I want `it` to point to `A`?

```cpp
auto it=ranges::find(
	tc::transform(v,std::mem_fn(&A::id)),
	4
).base();
```

---

#Transform Adaptor Implementation

```cpp
template<typename Base, typename Func>
struct transform_range {
	struct iterator {
	private:
		Func m_func; // in every iterator, hmmm...
		decltype( tc::begin(std::declval<Base&>()) ) m_it;
	public:
		decltype(auto) operator*() const {
			return m_func(*m_it);
		}
		decltype(auto) base() const {
			return (m_it);
		}
		...
	};
};
```

---

#Filter Adaptor

Range of all `a` with `a.id==4`?

```cpp
tc::filter(
	v,
	[](A const& a){ return 4==a.id; }
);
```

* Lazy! Filter executed while iterating

---

#Filter Adaptor Implementation

```cpp
template<typename Base, typename Func>
struct filter_range {
	struct iterator {
	private:
		Func m_func; // functor and TWO iterators!
		decltype( ranges::begin(std::declval<Base&>()) ) m_it;
		decltype( ranges::begin(std::declval<Base&>()) ) m_itEnd;
	public:
		iterator& operator++() {
			++m_it;
			while( m_it!=m_itEnd
				&& !static_cast<bool>(m_func(*m_it)) ) ++m_it;
					// why static_cast<bool> ?
			return *this;
		}
		...
	};
};
```

---

#How does iterator look like of

`tc::filter(tc::filter(tc::filter(...)))` ?

---

```cpp
m_func3
m_it3
	m_func2
	m_it2
		m_func1
		m_it1;
		m_itEnd1;
	m_itEnd2
		m_func1
		m_itEnd1;
		m_itEnd1;
m_itEnd3
	m_func2
	m_it2
		m_func1
		m_itEnd1;
		m_itEnd1;
	m_itEnd2
		m_func1
		m_itEnd1;
		m_itEnd1;
```
Boost.Range does this! ARGH!

---

#More Efficient Range Adaptors

Must keep iterators small

Idea:
adaptor object carries everything that is common for all iterators

```cpp
 m_func
 m_itEnd
```

Iterators carry reference to adaptor object (for common stuff) and base iterator
```cpp
 *m_rng
 m_it
```

- Iterator cannot outlive its range

---

#Again: How does iterator look like of

```cpp
tc::filter(tc::filter(tc::filter(...))) ?
```

```cpp
m_rng3
m_it3
	m_rng2
	m_it2
		m_rng1
		m_it1
```

- Range V3 State of The Art
- Still not insanely great...

---

#Index Concept

Index
- Like iterator
- But all operations require its range object

```cpp
template<typename Base, typename Func>
struct index_range {
	...
	using Index=...;
	Index begin_index() const;
	Index end_index() const;
	void increment_index( Index& idx ) const;
	void decrement_index( Index& idx ) const;
	reference dereference( Index& idx ) const;
	...
};
```

---

#Index-Iterator Compatibility

- Index from Iterator
  * `using Index = Iterator`
  * Index operations = Iterator operations

- Iterator from Index

```cpp
template<typename IndexRng>
struct iterator_for_index {
	IndexRng* m_rng
	typename IndexRng::Index m_idx;

iterator& operator++() {
		m_rng.increment_index(m_idx);
		return *this;
	}
	...
};
```

---

#Super-Efficient Range Adaptors With Indices

Index-based filter_range

```cpp
template<typename Base, typename Func>
struct filter_range {
	Func m_func;
	Base& m_base;

using Index=typename Base::Index;
	void increment_index( Index& idx ) const {
		do {
			m_base.increment_index(idx);
		} while( idx!=m_base.end_index()
			&& !m_func(m_base.dereference_index(idx))
		);
	}
};
```

---

#Super-Efficient Range Adaptors With Indices

Index-based filter_range

```cpp
template<typename Base, typename Func>
struct filter_range {
	Func m_func;
	Base& m_base;

using Index=typename Base::Index;
...
```

```cpp
template<typename IndexRng>
struct iterator_for_index {
	IndexRng* m_rng
	typename IndexRng::Index m_idx;
	...
```

- All iterators are two pointers
  * irrespective of stacking depth

---

#Range V3 and rvalue containers

If adaptor input is lvalue container
- `view::filter` creates view
- view is reference, O(1) copy, shallow constness etc.
  
```cpp
auto rng=view::filter(vec,pred1);
bool b=ranges::any_of(rng,pred2);
```

---

#Range V3 and rvalue containers

If adaptor input is rvalue container
- `view::filter` cannot create view
- view would hold dangling reference to rvalue
  
```cpp
auto rng=view::filter(create_vector(),pred1); // DOES NOT COMPILE
bool b=ranges::any_of(rng,pred2);
```

---

#Range V3 and rvalue containers

If adaptor input is rvalue container
- `view::filter` cannot create view
- view would hold dangling reference to rvalue
  
```cpp
auto rng=action::filter(create_vector(),pred1); // COMPILES
bool b=ranges::any_of(rng,pred2);
```

Big Trap
- not lazy anymore!

---

#think-cell and rvalue containers

If adaptor input is lvalue container
- `tc::filter` creates view
- view is reference, O(1) copy, shallow constness etc.

If adaptor input is rvalue container
- `tc::filter` creates container
- aggregates rvalue container, deep copy, deep constness etc.

Always lazy
- Laziness and container-ness are orthogonal concepts

```cpp
auto rng=tc::filter(vec,pred1);
bool b=ranges::any_of(rng,pred2);
```

```cpp
auto rng=tc::filter(creates_vector(),pred1);
bool b=ranges::any_of(rng,pred2);
```

---

##Beyond Range V3:
#More Flexible Algorithm Returns

```cpp
template< typename Rng, typename What >
decltype(auto) find( Rng && rng, What const& what ) {
	auto const itEnd=ranges::end(rng);
	for( auto it=ranges::begin(rng); it!=itEnd; ++it )
		if( *it==what )
			return it;
	return itEnd;
}
```

---

#More Flexible Algorithm Returns (2)

```cpp
struct return_iterator_or_end {
	static auto pack(auto it, auto&& rng) {
		return it;
	}
	static auto pack_singleton(auto && rng) {
		return ranges::end(rng);
	}
}
```
```cpp
auto it=find<return_iterator_or_end>(...)
```

---

#More Flexible Algorithm Returns (3)

```cpp
*struct return_iterator {
	static auto pack(auto it, auto&& rng) {
		return it;
	}
	static auto pack_singleton(auto && rng) {
*       std::assert(false);
		return ranges::end(rng);
	}
}
```

```cpp
auto it=find<return_iterator>(...)
```

---

#More Flexible Algorithm Returns (4)

```cpp
template< typename Pack, typename Rng, typename What >
decltype(auto) find( Rng && rng, What const& what ) {
	auto const itEnd=ranges::end(rng);
	for( auto it=ranges::begin(rng); it!=itEnd; ++it )
		if( *it==what )
			return Pack::pack(it,rng);
	return Pack::pack_singleton(rng);
}
```
```cpp
*struct return_head {
	static auto pack(auto it, auto&& rng) {
*       return tc::take( std::forward<decltype(rng)>(rng), it);
	}
	static auto pack_singleton(auto&& rng) {
		std::assert(false);
		return tc::take( std::forward<decltype(rng)>(rng), ranges::begin(rng));
	}
}
```
```cpp
auto it=find<return_head>(...)
```

---

#Generator Ranges

```cpp
template<typename Func>
void traverse_widgets( Func func ) {
	if( window1 ) {
		window1->traverse_widgets(std::ref(func));
	}
	func(button1);
	func(listbox1);
	if( window2 ) {
		window2->traverse_widgets(std::ref(func));
	}
}
```

- like range of widgets
- but no iterators

---

#Generator Ranges

```cpp
mouse_hit_any_widget=tc::any_of(
	[](auto func){ traverse_widgets(func); },
	[](auto const& widget) {
		return widget.mouse_hit();
	}
);
```

---

#External Iteration

* Consumer calls producer to get new element
* example: C++ iterators

```
^
| Stack           Producer            Producer
|               /          \        /          \
		Consumer            Consumer            Consumer
```

- Consumer is at bottom of stack
- Producer is at top of stack

---

#External iteration (2)

Consumer is at bottom of stack
* contiguous code path for whole range
* easier to write
* better performance
  * state encoded in instruction pointer
  * no limit for stack memory

Producer is at top of stack
* contiguous code path for each item
* harder to write
* worse performance
  * single entry point, must restore state
  * fixed amount of memory or go to heap

---
  
#Internal Iteration
* Producer calls consumer to offer new element
* example: EnumThreadWindows

```
^
| Stack           Consumer            Consumer
|               /          \        /          \
		Producer            Producer            Producer
```

Producer is at bottom of stack
* ... all the advantages of being bottom of stack ...

Consumer is at top of stack
* ... all the disadvantages of being top of stack ...

---

#Coroutines

Can both consumer and producer be bottom-of-stack?
* Yes, with coroutines

```cpp
// does not compile, conceptual
generator<widget&> traverse_widgets() {
	if( window1 ) {
		window1->traverse_widgets();
	}
	yield button1;
	yield listbox1;
	if( window2 ) {
		window2->traverse_widgets();
	}
}
```

---

#Coroutines (2)

```cpp
// does not compile, conceptual
generator<widget&> traverse_widgets() {
	if( window1 ) {
*		window1->traverse_widgets();
	}
	yield button1;
	yield listbox1;
	if( window2 ) {
*		window2->traverse_widgets();
	}
}
```

---

#Coroutines (2)

- Stackful
  * use two stacks and switch between them
  * very expensive
      - implemented as OS fibers
      - 1 MB of virtual memory per coroutine
- Stackless (N4402)
  * can only yield in top-most function
  * still a bit expensive
      - dynamic jump to resume point
      - save/restore some registers
      - no aggressive inlining

---

#Internal Iteration often good enough

Algorithm               | Internal Iteration?
----------------------- | -------------------
find                    | no (single pass iterators)
binary_search           | no (random access iterators)
--

Adaptor                 | Internal Iteration?
----------------------- | -------------------
tc::filter              | yes
tc::transform           | yes

So allow ranges that support only internal iteration!

---

#any_of implementation

```cpp
namespace tc {
	template< typename Rng >
	bool any_of( Rng const& rng ) {
		bool bResult=false;
		tc::enumerate( rng, [&](bool_context b){
			bResult=bResult || b;
		} );
		return bResult;
	}
}
```

- `tc::enumerate` is common interface for iterator, index and generator ranges

- Ok?
--

* std::any_of stops when true is encountered!

---

#Interruptable Generator Ranges

First idea: exception!

- too slow:-(

Second idea:
```cpp
enum break_or_continue {
	break_,
	continue_
};
```

```cpp
template< typename Rng >
bool any_of( Rng const& rng ) {
	bool bResult=false;
	tc::enumerate( rng, [&](bool_context b){
		bResult=bResult || b;
		return bResult ? break_ : continue_;
	} );
	return bResult;
}
```

---

#Interruptable Generator Ranges (2)

- Generator Range can elide `break_` check
	* If functor returns `break_or_continue`,
		 * break if `break_` is returned.
	* If functor returns anything else,
		* nothing to check, always continue
		
---

#concat

```cpp
std::list<int> lst;
std::vector<int> vec;

std::for_each( tc::concat(lst,vec), [](int i) {
	...
});
```

---

#concat implementation with indices

```cpp
template<typename Rng1, typename Rng2>
struct concat_range {
private:
	using Index1=typename range_index<Rng1>::type;
	using Index2=typename range_index<Rng2>::type;

Rng1& m_rng1;
	Rng2& m_rng2;
	using index=std::variant<Index1, Index2>;
public:
...
```

---

#concat implementation with indices (2)

```cpp
...
	void increment_index(index& idx) {
		idx.switch(
			[&](Index1& idx1){
				m_rng1.increment_index(idx1);
				if (m_rng1.at_end_index(idx1)) {
					idx=m_rng2.begin_index();
				}
			},
			[&](Index2& idx2){
				m_rng2.increment_index(idx2);
			}
		);
	}
...
```
- Branch for each increment!

---

#concat implementation with indices (3)

```cpp
...
	auto dereference_index(index const& idx) const {
		return idx.switch(
			[&](Index1 const& idx1){
				return m_rng1.dereference(idx1);
			},
			[&](Index2 const& idx2){
				return m_rng2.dereference(idx2);
			}
		);
	}
	...
};
```

- Branch for each dereference!
- How avoid all these branches?
--

* With Generator Ranges!

---

#concat implementation as generator range

```cpp
template<typename Rng1, typename Rng2>
struct concat_range {
private:
	Rng1 m_rng1;
	Rng2 m_rng2;

public:

...

// version for non-breaking func
	template<typename Func>
	void operator()(Func func) {
		tc::fn_enumerate(m_rng1, func);
		tc::fn_enumerate(m_rng2, func);
	}
};
```

- Even iterator-based ranges sometimes perform better with generator interface!

---

Now that we have all this range stuff

- URL of our range library: 
	https://github.com/think-cell/think-cell-library
	
--

#I hate the range-based for loop!

because it encourages people to write this

```cpp
bool b=false;
for( int n : rng ) {
	if( is_prime(n) ) {
		b=true;
		break;
	}
}
```

instead of this

```cpp
bool b=tc::any_of( rng, is_prime );
```

THANK YOU!

共有する