[Bug ipa/115097] Strange suboptimal codegen specifically at -O2 when copying struct type

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115097

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |ipa
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
So actually it seems that the reason is ICF plus inlining:

t.ii:2:3: optimized: Semantic equality hit:A test1(A&)/0->A test2(A&&)/1
t.ii:2:3: optimized: Assembler symbol names:_Z5test1R1A/0->_Z5test2O1A/1
t.ii:2:3: optimized: Semantic equality hit:A test1(A&)/0->A test3(const A&)/2
t.ii:2:3: optimized: Assembler symbol names:_Z5test1R1A/0->_Z5test3RK1A/2
t.ii:2:3: optimized: Semantic equality hit:A test1(A&)/0->A test4(const A&&)/3
t.ii:2:3: optimized: Assembler symbol names:_Z5test1R1A/0->_Z5test4OK1A/3
optimized:  Inlined A test1(A&)/4 into A test2(A&&)/1 which now has time
4.000000 and size 5, net change of -1.
optimized:  Inlined A test1(A&)/5 into A test3(const A&)/2 which now has time
4.000000 and size 5, net change of -1.
optimized:  Inlined A test1(A&)/6 into A test4(const A&&)/3 which now has time
4.000000 and size 5, net change of -1.

for some reason we "optimize" the functions to the following in IPA ICF:

struct A test4 (const struct A & a)
{
  struct A retval.6;

  <bb 2> [local count: 1073741824]:
  retval.6 = test1 (a_2(D)); [tail call]
  return retval.6;

}


struct A test3 (const struct A & a)
{
  struct A retval.5;

  <bb 2> [local count: 1073741824]:
  retval.5 = test1 (a_2(D)); [tail call]
  return retval.5;

}


struct A test2 (struct A & a)
{
  struct A retval.4;

  <bb 2> [local count: 1073741824]:
  retval.4 = test1 (a_2(D)); [tail call]
  return retval.4;

}

and then we inline them back, introducing the extra copy.  Why do we use
tail-calls here instead of aliases?  Why do we lack cost modeling here?
Why do we inline back?  It looks like a pointless exercise to me ...
With -fdisable-ipa-inline we get

_Z5test2O1A:
.LFB5:
        .cfi_startproc
        jmp     _Z5test1R1A

so that's at least reasonable and what's expected I suppose.  So one
could argue the bug is in the inliner and with introducing the extra
copy (IIRC there's a bug about this), but still.