[Bug tree-optimization/90883] Generated code is worse if returned struct is unnamed

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90883

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org,
                   |                            |law at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
          Component|c++                         |tree-optimization

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So before SRA we have

slow ()
{
  struct C D.26322;
  struct C D.29886;

  <bb 2> :
  D.26322 = {};
  D.26322.a = {};
  D.26322.b = 0;
  D.29886 = D.26322;
  D.26322 ={v} {CLOBBER};
  return D.29886;

fast ()
{
  struct C c;
  struct C D.29889;

  <bb 2> :
  c = {};
  D.29889 = c;
  c ={v} {CLOBBER};
  return D.29889;

and then SRA does nothing to fast but half-way optimizes slow() to

  <bb 2> :
  D.26322 = {};
  D.26322.a = {};
  D.29886 = D.26322;
  D.26322 ={v} {CLOBBER};
  return D.29886;

where then probably DSE wrecks things by trimming the earlier store:

  <bb 2> :
  MEM[(struct C *)&D.26322 + 7B] = {};
  D.26322.a = {};
  D.29886 = D.26322;
  D.26322 ={v} {CLOBBER};
  return D.29886;

which we later do not recover from (store-merging?).  I also wonder
why SRA does not elide the aggregate copy.

-O2 -fno-tree-dse generates

_Z4slowv:
.LFB1099:
        .cfi_startproc
        movq    $0, -32(%rsp)
        xorl    %eax, %eax
        xorl    %edx, %edx
        ret