[Bug middle-end/113907] [14 regression] ICU miscompiled since on x86 since r14-5109-ga291237b628f41

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113907

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #20 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Ah, I think it is IPA-ICF.
What happens is that fnsplit splits the uprv_copyArray{16,32,64} functions,
where the
inline part does what is actually different among the functions, i.e. the
tests which among other do the early out for
length<0 || (length&1)!=0
or
length<0 || (length&3)!=0
or
length<0 || (length&7)!=0
among other things, and then the *.part.0 parts which are exactly the same for
the 3 functions IL wise, except for the unfortunately differences in the
ranges.
So, essentially just
  <bb 2> [local count: 132713219]:
  _2 = length_1(D) != 0;
  _5 = inData_3(D) != outData_4(D);
  _6 = _2 & _5;
  if (_6 != 0)
    goto <bb 3>; [33.00%]
  else
    goto <bb 4>; [67.00%]

  <bb 3> [local count: 43795362]:
  # RANGE [irange] unsigned int [N, 2147483647] MASK 0x7ffffffe VALUE 0x0
  length.43_7 = (unsigned int) length_1(D);
  # USE = anything
  # CLB = anything
  memcpy (outData_4(D), inData_3(D), length.43_7);

  <bb 4> [local count: 132713219]:

  <bb 5> [local count: 132713219]:
  # RANGE [irange] int [0, 0][N, +INF] MASK 0x7fffffff VALUE 0x0
  # _8 = PHI <length_1(D)(4)>
  return _8;
where N is 2, 4 or 8.
Now IPA-ICF comes, doesn't care about ranges, merges all 3 into one and picks
there
the most unfortunate case, the case from the copyData64 part with N 8.
Later on we inline this back into the functions, the function splitting was
useless.

So, the question is what to do in IPA-ICF.  Should we punt on range
differences, reset all ranges or try to merge the ranges from all the functions
merged together?
I think the last would be best.