<vector>: vector's move ctor initializes members twice

[StephanTLavavej]

@xiangfan-ms discovered that vector's move constructor zero-initializes its pointers before taking the other vector's pointers, and that this isn't optimized away in release mode. (Which results in larger, slower codegen, and probably inhibits further optimizations like memcpy coalescing.)

1. vector's move constructor constructs its base before allocating the proxy and taking the other's pointers:
   

   STL/stl/inc/vector

   Lines 581 to 584 in 530bdc5

   	vector(vector&& _Right) noexcept : _Mypair(_One_then_variadic_args_t{}, _STD move(_Right._Getal())) {
   	_Mypair._Myval2._Alloc_proxy(_GET_PROXY_ALLOCATOR(_Alty, _Getal()));
   	_Move_construct(_Right, true_type{});
   	}

2. The base construction sets the pointers to null:
   

   STL/stl/inc/vector

   Line 375 in 530bdc5

   _Vector_val() noexcept : _Myfirst(), _Mylast(), _Myend() {}

3. Then _Move_construct calls _Take_contents which re-assigns the pointers:
   

   STL/stl/inc/vector

   Lines 556 to 558 in 530bdc5

   	void _Move_construct(vector& _Right, true_type) noexcept { // move from _Right, stealing its contents
   	_Mypair._Myval2._Take_contents(_Right._Mypair._Myval2);
   	}

   
   

   STL/stl/inc/vector

   Lines 384 to 393 in 530bdc5

   	void _Take_contents(_Vector_val& _Right) noexcept {
   	this->_Swap_proxy_and_iterators(_Right);
   	_Myfirst = _Right._Myfirst;
   	_Mylast = _Right._Mylast;
   	_Myend = _Right._Myend;
   	_Right._Myfirst = pointer();
   	_Right._Mylast = pointer();
   	_Right._Myend = pointer();
   	}

I believe it's possible to change this behavior. We'd need to make _Vector_val constructible from three pointers, so we can directly initialize them. (This will work with _Compressed_pair's one-then-variadic-args - the one arg is for the allocator, the variadic args are sent to _Vector_val.) Then we'd need to allocate/swap proxies (which vanishes for release mode), and null out the other pointers. (This is noexcept so allocation failure is termination - no EH concerns. vNext will remove the allocation.)

[BillyONeal]

string has similar cases we didn't touch we because we were being paranoid about ABI compat at the time.

[StephanTLavavej]

I think we've gotten more comfortable about the ABI implications of changing the behavior of single member functions in a way that continues to uphold their Standard-mandated behavior. I see no potential for mix-and-match problems here.

[CaseyCarter]

I read Billy's comment less as a warning and more as a suggestion that we might want to make similar changes to basic_string.

[BillyONeal]

I read Billy's comment less as a warning and more as a suggestion that we might want to make similar changes to basic_string.

Yes, that's right. I meant in general that this problem where we have bases that initialize things happens in string too.

(Of course in basic_string we might have to add a no-op ctor that takes a tag to give it a different name or something to _String_val if there is ABI paranoia but still...)