[Bug tree-optimization/90020] [7/8/9 regression] -O2 -Os x86-64 wrong code generated for GNU Emacs

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90020

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
                 CC|                            |rguenth at gcc dot gnu.org
            Summary|[8 regression] -O2 -Os      |[7/8/9 regression] -O2 -Os
                   |x86-64 wrong code generated |x86-64 wrong code generated
                   |for GNU Emacs               |for GNU Emacs
      Known to fail|                            |7.4.0, 8.3.0, 9.0

--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> ---
I can confirm that even though the code is not so nice :)
I have 2 versions of the reduced test-case:

1)
$ gcc emacs0.c -Os -fno-strict-aliasing -g && valgrind ./a.out
==20746== Memcheck, a memory error detector
==20746== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==20746== Using Valgrind-3.14.0 and LibVEX; rerun with -h for copyright info
==20746== Command: ./a.out
==20746== 
==20746== Invalid read of size 8
==20746==    at 0x4011CF: select_window (emacs0.c:116)
==20746==    by 0x40105A: main (emacs0.c:126)
==20746==  Address 0xa is not stack'd, malloc'd or (recently) free'd

fails for 4.8.0+ except 4.9.x releases.

2)
$ gcc emacs.c -Os -fno-strict-aliasing && ./a.out

started failing with r238242

The problematic transformation:

  <bb 2> [local count: 1073741824]:
  _1 = PSEUDOVECTORP (window_6(D));
  pretmp_9 = MEM[(struct window *)window_6(D)].contents;
  if (_1 != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 4>; [50.00%]

happens in PRE. Richi can you please take a look?