From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-bugzilla@gcc.gnu.org>
Received: by sourceware.org (Postfix, from userid 48)
 id A7F863858C52; Fri, 12 Aug 2022 13:47:37 +0000 (GMT)
DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org A7F863858C52
From: "jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug rtl-optimization/106590] [12/13 Regression] x86-64
 miscompilation starting with r12-8233-g1ceddd7497e15d w/ mtune=skylake
Date: Fri, 12 Aug 2022 13:47:37 +0000
X-Bugzilla-Reason: CC
X-Bugzilla-Type: changed
X-Bugzilla-Watch-Reason: None
X-Bugzilla-Product: gcc
X-Bugzilla-Component: rtl-optimization
X-Bugzilla-Version: 12.1.0
X-Bugzilla-Keywords: wrong-code
X-Bugzilla-Severity: normal
X-Bugzilla-Who: jakub at gcc dot gnu.org
X-Bugzilla-Status: NEW
X-Bugzilla-Resolution: 
X-Bugzilla-Priority: P2
X-Bugzilla-Assigned-To: unassigned at gcc dot gnu.org
X-Bugzilla-Target-Milestone: 12.2
X-Bugzilla-Flags: 
X-Bugzilla-Changed-Fields: 
Message-ID: <bug-106590-4-KNxcwpIAKx@http.gcc.gnu.org/bugzilla/>
In-Reply-To: <bug-106590-4@http.gcc.gnu.org/bugzilla/>
References: <bug-106590-4@http.gcc.gnu.org/bugzilla/>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
X-Bugzilla-URL: http://gcc.gnu.org/bugzilla/
Auto-Submitted: auto-generated
MIME-Version: 1.0
X-BeenThere: gcc-bugs@gcc.gnu.org
X-Mailman-Version: 2.1.29
Precedence: list
List-Id: Gcc-bugs mailing list <gcc-bugs.gcc.gnu.org>
List-Unsubscribe: <https://gcc.gnu.org/mailman/options/gcc-bugs>,
 <mailto:gcc-bugs-request@gcc.gnu.org?subject=unsubscribe>
List-Archive: <https://gcc.gnu.org/pipermail/gcc-bugs/>
List-Post: <mailto:gcc-bugs@gcc.gnu.org>
List-Help: <mailto:gcc-bugs-request@gcc.gnu.org?subject=help>
List-Subscribe: <https://gcc.gnu.org/mailman/listinfo/gcc-bugs>,
 <mailto:gcc-bugs-request@gcc.gnu.org?subject=subscribe>
X-List-Received-Date: Fri, 12 Aug 2022 13:47:37 -0000

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D106590

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Cleaned up testcase:

/* { dg-additional-options "-mtune=3Dskylake" { target { i?86-*-* x86_64-*-=
* } }
} */

typedef struct A { short a; } A;
typedef A *B;
typedef struct C { int c, d; } C;
typedef C *D;

B
foo (void)
{
  static A r =3D { .a =3D 1 };
  return &r;
}

D
bar (void)
{
  static C r =3D { .c =3D 1, .d =3D 23 };
  return &r;
}

static inline int __attribute__((always_inline))
baz (short a)
{
  int e =3D 1, f;
  short g;
  D h;

  switch (a)
    {
    case 1:
      f =3D 23;
      g =3D 1;
      break;
    case 2:
      f =3D 20;
      g =3D 2;
      break;
    }

  h =3D bar ();

  if (h->d !=3D f || h->c !=3D g)
    __builtin_abort ();
  return e;
}

int
qux (void)
{
  B i =3D foo ();
  int e =3D 1;

  switch (i->a)
    {
    case 1:
    case 2:
      e =3D baz (i->a);
      break;
    case 3:
      e =3D 0;
      break;
    }

  return e;
}

int
main ()
{
  qux ();
  return 0;
}

I think this is a latent bug in noce_convert_multiple_sets_1.
For cond
(eq (reg/v:HI 82 [ g ])
    (const_int 1 [0x1]))
and cc_cmp
(eq (reg:CCZ 17 flags)
    (const_int 0 [0]))
when we try_emit_cmove_seq for
(insn 3 35 4 7 (set (reg/v:HI 82 [ g ])
        (const_int 2 [0x2])) "pr106590.c":37:9 84 {*movhi_internal}
     (nil))
insn, we get seq1
(insn 65 0 66 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg/v:HI 82 [ g ])
            (const_int 1 [0x1]))) -1
     (nil))

(insn 66 65 67 (set (reg:QI 96)
        (ne:QI (reg:CCZ 17 flags)
            (const_int 0 [0]))) -1
     (nil))

(insn 67 66 68 (set (reg:HI 95)
        (zero_extend:HI (reg:QI 96))) -1
     (nil))

(insn 68 67 0 (parallel [
            (set (reg:HI 95)
                (plus:HI (reg:HI 95)
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) -1
     (nil))
and seq2
(insn 69 0 70 (set (reg:HI 97)
        (const_int 2 [0x2])) -1
     (nil))

(insn 70 69 0 (set (reg:HI 95)
        (if_then_else:HI (eq (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (reg/v:HI 82 [ g ])
            (reg:HI 97))) -1
     (nil))
and later for
(insn 4 3 36 7 (set (reg/v:SI 84 [ f ])
        (const_int 20 [0x14])) "pr106590.c":36:9 83 {*movsi_internal}
     (nil))
we get seq1
(insn 72 0 71 (set (reg:SI 98)
        (const_int 20 [0x14])) -1
     (nil))

(insn 71 72 73 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg/v:HI 82 [ g ])
            (const_int 1 [0x1]))) -1
     (nil))

(insn 73 71 0 (set (reg/v:SI 84 [ f ])
        (if_then_else:SI (eq (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (reg/v:SI 84 [ f ])
            (reg:SI 98))) -1
     (nil))
and seq2
(insn 74 0 75 (set (reg:SI 99)
        (const_int 20 [0x14])) -1
     (nil))

(insn 75 74 0 (set (reg/v:SI 84 [ f ])
        (if_then_else:SI (eq (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (reg/v:SI 84 [ f ])
            (reg:SI 99))) -1
     (nil))
In the former case, cost1 =3D=3D cost2, in the latter case cost1 > cost2.
So, we for the first insn choose seq1 and for the second seq2.
That is wrong, because the first seq1 clobbers the (reg 17 flags) register
which is used in cc_cmp/rev_cc_cmp.  So, if we pick that seq1, we may not u=
se
any seq2 afterwards for the latter instructions, need to always select seq1
since then.=