[Bug libstdc++/110287] _M_check_len is expensive

["hubicka at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110287

--- Comment #9 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
This is _M_realloc insert at release_ssa time:

eleased 63 names, 165.79%, removed 63 holes
void std::vector<pair_t>::_M_realloc_insert<const pair_t&> (struct vector *
const this, struct iterator __position, const struct pair_t & __args#0)
{ 
  struct pair_t * const __position;
  struct pair_t * __new_finish;
  struct pair_t * __old_finish;
  struct pair_t * __old_start;
  long unsigned int _1; 
  struct pair_t * _2;
  struct pair_t * _3;
  long int _4;
  long unsigned int _5;
  struct pair_t * _6;
  const size_type _10;
  long int _13;
  struct pair_t * iftmp.5_15;
  struct pair_t * _17; 
  struct _Vector_impl * _18;
  long unsigned int _22;
  long int _23;
  long unsigned int _24;
  long unsigned int _25;
  struct pair_t * _26;
  long unsigned int _36;

  <bb 2> [local count: 1073741824]:
  __position_27 = MEM[(struct __normal_iterator *)&__position];
  _10 = std::vector<pair_t>::_M_check_len (this_8(D), 1,
"vector::_M_realloc_insert");
  __old_start_11 = this_8(D)->D.25975._M_impl.D.25282._M_start;
  __old_finish_12 = this_8(D)->D.25975._M_impl.D.25282._M_finish;
  _13 = __position_27 - __old_start_11;
  if (_10 != 0)
    goto <bb 3>; [54.67%]
  else
    goto <bb 4>; [45.33%]

  <bb 3> [local count: 587014656]:
  _18 = &MEM[(struct _Vector_base *)this_8(D)]._M_impl;
  _17 = std::__new_allocator<pair_t>::allocate (_18, _10, 0B);

  <bb 4> [local count: 1073741824]:
  # iftmp.5_15 = PHI <0B(2), _17(3)>
  _1 = (long unsigned int) _13;
  _2 = iftmp.5_15 + _1;
  *_2 = *__args#0_14(D);
  if (_13 > 0)
    goto <bb 5>; [41.48%]
  else
    goto <bb 6>; [58.52%]

  <bb 5> [local count: 445388112]:
  __builtin_memmove (iftmp.5_15, __old_start_11, _1);

  <bb 6> [local count: 1073741824]:
  _36 = _1 + 8;
  __new_finish_16 = iftmp.5_15 + _36;
  _23 = __old_finish_12 - __position_27;
  if (_23 > 0)
    goto <bb 7>; [41.48%]
  else
    goto <bb 8>; [58.52%]

  <bb 7> [local count: 445388112]:
  _24 = (long unsigned int) _23;
  __builtin_memcpy (__new_finish_16, __position_27, _24);

  <bb 8> [local count: 1073741824]:
  _25 = (long unsigned int) _23;
  _26 = __new_finish_16 + _25;
  _3 = this_8(D)->D.25975._M_impl.D.25282._M_end_of_storage;
  _4 = _3 - __old_start_11;
  if (__old_start_11 != 0B)
    goto <bb 9>; [53.47%]
  else
    goto <bb 10>; [46.53%]

  <bb 9> [local count: 574129752]:
  _22 = (long unsigned int) _4;
  operator delete (__old_start_11, _22);

  <bb 10> [local count: 1073741824]:
  this_8(D)->D.25975._M_impl.D.25282._M_start = iftmp.5_15;
  this_8(D)->D.25975._M_impl.D.25282._M_finish = _26;
  _5 = _10 * 8;
  _6 = iftmp.5_15 + _5;
  this_8(D)->D.25975._M_impl.D.25282._M_end_of_storage = _6;
  return;

}

First it is not clear to me why we need memmove at all?

So first issue is:
  <bb 2> [local count: 1073741824]:
  __position_27 = MEM[(struct __normal_iterator *)&__position];
  _10 = std::vector<pair_t>::_M_check_len (this_8(D), 1,
"vector::_M_realloc_insert");
  __old_start_11 = this_8(D)->D.25975._M_impl.D.25282._M_start;
  __old_finish_12 = this_8(D)->D.25975._M_impl.D.25282._M_finish;
  _13 = __position_27 - __old_start_11;
  if (_10 != 0)
    goto <bb 3>; [54.67%]
  else
    goto <bb 4>; [45.33%]

Without inlining _M_check_len early we can not work out return value range,
since we need to know that paramter 2 is 1 and not 0.
Adding __builtin_unreachable check after helps to reduce

if (_10 != 0)

but I need to do something about inliner accounting the conditional to function
body size.

  <bb 4> [local count: 1073741824]:
  # iftmp.5_15 = PHI <0B(2), _17(3)>
  _1 = (long unsigned int) _13;
  _2 = iftmp.5_15 + _1;
  *_2 = *__args#0_14(D);
  if (_13 > 0)
    goto <bb 5>; [41.48%]
  else
    goto <bb 6>; [58.52%]

  <bb 5> [local count: 445388112]:
  __builtin_memmove (iftmp.5_15, __old_start_11, _1);

Is this code about inserting value to the middle?  Since push_back always
initializes iterator to point to the end, this seems quite sily to do.
Can't we do somehting like _M_realloc_append?