[Bug tree-optimization/58662] New: wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[Bug tree-optimization/58662] New: wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[su at cs dot ucdavis.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

            Bug ID: 58662
           Summary: wrong code at -O2 and -O3 on x86_64-linux-gnu (in
                    64-bit mode)
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: su at cs dot ucdavis.edu

The current gcc trunk miscompiles the attached testcase on x86_64-linux-gnu at
-O2 and -O3 in 64-bit mode. 

It is a regression from 4.8.x.  


$ gcc-trunk -v
Using built-in specs.
COLLECT_GCC=gcc-trunk
COLLECT_LTO_WRAPPER=/usr/local/gcc-trunk/libexec/gcc/x86_64-unknown-linux-gnu/4.9.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: ../gcc-trunk/configure
--enable-languages=c,c++,objc,obj-c++,fortran,lto --disable-werror
--enable-checking=release --with-gmp=/usr/local/gcc-trunk
--with-mpfr=/usr/local/gcc-trunk --with-mpc=/usr/local/gcc-trunk
--with-cloog=/usr/local/gcc-trunk --prefix=/usr/local/gcc-trunk
Thread model: posix
gcc version 4.9.0 20131007 (experimental) [trunk revision 203235] (GCC) 
$ 
$ gcc-trunk -Os small.c; a.out
0
$ gcc-trunk -O2 small.c; a.out
-1008071096
$ gcc-trunk -O3 small.c; a.out
86090104
$
$ gcc-4.8 -O2 small.c; a.out
0
$ gcc-4.8 -O3 small.c; a.out
0 
$ 

-------------------------------------

int printf (const char *, ...);

int a, b, c, d;

int
foo (int p1, short p2)
{
  return p1 / p2;
}

int
main ()
{
  char e;
  d = foo (a == 0, (0, 35536)); 
  e = d % 14;
  b = e && c;
  printf ("%d\n", b);
  return 0;
}

[Bug tree-optimization/58662] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[mpolacek at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

Marek Polacek <mpolacek at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-10-08
                 CC|                            |mpolacek at gcc dot gnu.org
   Target Milestone|---                         |4.9.0
     Ever confirmed|0                           |1

--- Comment #1 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
Confirmed.  It seems to be the VRP; adding -fno-tree-vrp makes trunk behave as
4.8.

[Bug tree-optimization/58662] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[mpolacek at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

Marek Polacek <mpolacek at gcc dot gnu.org> changed:

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

--- Comment #2 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
Started with r202944.

[Bug tree-optimization/58662] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[mpolacek at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

--- Comment #3 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
Actually, I think it's the uncprop:

--- Q.c.139t.crited2    2013-10-08 13:03:04.169955615 +0200
+++ Q.c.141t.uncprop1    2013-10-08 13:03:04.169955615 +0200
@@ -51,7 +51,7 @@
   _13 = (int) _8;

   <bb 4>:
-  # iftmp.3_1 = PHI <0(5), _13(3)>
+  # iftmp.3_1 = PHI <_7(5), _13(3)>
   b = iftmp.3_1;
   printf ("%d\n", iftmp.3_1);
   return 0;

so -fno-tree-dominator-opts makes the bug go away.

[Bug tree-optimization/58662] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I will have a looksee.

[Bug tree-optimization/58662] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
It looks like changing the testcase line

  d = foo (a == 0, (0, 35536)); 

to

  d = foo (a == 0, -30000);  

fixes it.  _No_ IL difference in 065t.mergephi2, but then 066t.vrp1 has

 main ()
 {
   char e;
@@ -39,6 +35,7 @@
   _Bool _4;
   int _5;
   int _7;
+  int c.4_9;
   int b.5_11;
   int _12;

@@ -46,22 +43,37 @@
   a.0_3 = a;
   _4 = a.0_3 == 0;
   _5 = (int) _4;
-  _12 = 0;
-  d = 0;
-  _7 = 0;
-  e_8 = 0;
-  iftmp.3_1 = 0;
-  b ={v} 0;
-  b.5_11 ={v} b;
-  if (b.5_11 != 0)
+  _12 = _5 / -30000;
+  d = _12;
+  _7 = _12 % 14;
+  e_8 = (char) _7;
+  if (_7 != 0)
     goto <bb 3>;
   else
     goto <bb 4>;

   <bb 3>:
-  abort ();
+  c.4_9 = c;
+  if (c.4_9 != 0)
+    goto <bb 5>;
+  else
+    goto <bb 4>;

   <bb 4>:
+
+  <bb 5>:
+  # iftmp.3_1 = PHI <1(3), 0(4)>
+  b ={v} iftmp.3_1;
+  b.5_11 ={v} b;
+  if (b.5_11 != 0)
+    goto <bb 6>;
+  else
+    goto <bb 7>;
+
+  <bb 6>:
+  abort ();
+
+  <bb 7>:
   return 0;

 }

what?!

Reduced testcase:

extern void abort (void);

int a, c, d;
volatile int b;

static int
foo (int p1, short p2)
{
  return p1 / p2;
}

int
main ()
{
  char e;
  d = foo (a == 0, (0, 35536)); 
  e = d % 14;
  b = e && c;
  if (b != 0)
    abort ();
  return 0;
}

[Bug rtl-optimization/58662] [4.9 Regression] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
          Component|tree-optimization           |rtl-optimization
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org
            Summary|wrong code at -O2 and -O3   |[4.9 Regression] wrong code
                   |on x86_64-linux-gnu (in     |at -O2 and -O3 on
                   |64-bit mode)                |x86_64-linux-gnu (in 64-bit
                   |                            |mode)

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Doesn't fail with -O2 -fno-tree-vrp.

The uncprop transform is correct.

  _7 = _12 % 14;
  if (_7 != 0)
    goto <bb 3>;
  else
    goto <bb 7>;

  <bb 7>:
  goto <bb 4>;

  <bb 3>:
  c.4_9 = c;
  _8 = c.4_9 != 0;
  _13 = (int) _8;

  <bb 4>:
  iftmp.3_1 = PHI <_7(7), _13(3)>

we know that _7 is zero on the edge from BB 7.

What we now do is coalesce iftmp.3_1 with both _7 and _13 with also seems
correct.  The initial RTL generated looks ok as well:

(insn 25 24 26 2 (parallel [
            (set (reg:SI 83 [ D.1762 ])
                (minus:SI (reg:SI 90 [ D.1762 ])
                    (reg:SI 106)))
            (clobber (reg:CC 17 flags))
        ]) t.c:17 296 {*subsi_1}
     (nil))
(insn 26 25 27 2 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg:SI 83 [ D.1762 ])
            (const_int 0 [0]))) t.c:18 3 {*cmpsi_ccno_1}
     (nil))
(jump_insn 27 26 28 2 (set (pc)
        (if_then_else (eq (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (label_ref 33)
            (pc))) t.c:18 616 {*jcc_1}
     (int_list:REG_BR_PROB 5000 (nil))
 -> 33)

;;  succ:       4 [50.0%]  (FALLTHRU)
;;              5 [50.0%]

;; basic block 4, loop depth 0, count 0, freq 5000
;;  prev block 2, next block 5, flags: (NEW, REACHABLE, RTL)
;;  pred:       2 [50.0%]  (FALLTHRU)
...
(insn 32 31 33 4 (set (reg:SI 83 [ D.1762 ])
        (reg:SI 107 [ D.1763 ])) t.c:18 86 {*movsi_internal}
     (nil))
;;  succ:       5 [100.0%]  (FALLTHRU)

;; basic block 5, loop depth 0, count 0, freq 10000, maybe hot
;;  prev block 4, next block 6, flags: (NEW, REACHABLE, RTL)
;;  pred:       2 [50.0%]
;;              4 [100.0%]  (FALLTHRU)
(code_label 33 32 34 5 2 "" [1 uses])
(note 34 33 35 5 [bb 5] NOTE_INSN_BASIC_BLOCK)
(insn 35 34 36 5 (set (mem/v/c:SI (symbol_ref:DI ("b")  <var_decl
0x7fc8088581c8 b>) [2 b+0 S4 A32])
        (reg:SI 83 [ D.1762 ])) t.c:18 86 {*movsi_internal}
     (nil))

it is combine where mysteriously all defining instructions for reg:SI 83
vanish.  Not mine.

[Bug rtl-optimization/58662] [4.9 Regression] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> It looks like changing the testcase line
> 
>   d = foo (a == 0, (0, 35536)); 
> 
> to
> 
>   d = foo (a == 0, -30000);  
> 
> fixes it.  _No_ IL difference in 065t.mergephi2, but then 066t.vrp1 has
> 
>  main ()
>  {
>    char e;
> @@ -39,6 +35,7 @@
>    _Bool _4;
>    int _5;
>    int _7;
> +  int c.4_9;
>    int b.5_11;
>    int _12;
>  
> @@ -46,22 +43,37 @@
>    a.0_3 = a;
>    _4 = a.0_3 == 0;
>    _5 = (int) _4;
> -  _12 = 0;
> -  d = 0;
> -  _7 = 0;
> -  e_8 = 0;
> -  iftmp.3_1 = 0;
> -  b ={v} 0;
> -  b.5_11 ={v} b;
> -  if (b.5_11 != 0)
> +  _12 = _5 / -30000;
> +  d = _12;
> +  _7 = _12 % 14;
> +  e_8 = (char) _7;
> +  if (_7 != 0)
>      goto <bb 3>;
>    else
>      goto <bb 4>;
>  
>    <bb 3>:
> -  abort ();
> +  c.4_9 = c;
> +  if (c.4_9 != 0)
> +    goto <bb 5>;
> +  else
> +    goto <bb 4>;
>  
>    <bb 4>:
> +
> +  <bb 5>:
> +  # iftmp.3_1 = PHI <1(3), 0(4)>
> +  b ={v} iftmp.3_1;
> +  b.5_11 ={v} b;
> +  if (b.5_11 != 0)
> +    goto <bb 6>;
> +  else
> +    goto <bb 7>;
> +
> +  <bb 6>:
> +  abort ();
> +
> +  <bb 7>:
>    return 0;
>  
>  }
> 
> what?!

That's because there _is_ a IL difference.  With (0, 35536) the
-30000 tree constant has TREE_OVERFLOW set while with literal
-30000 in the source it does not.

Due to VRP relying on some fold-const.c overflow detection with
TREE_OVERFLOW this effect is to be expected.

[but in reality GIMPLE should not have any TREE_OVERFLOW constants
in the IL, though that's hard to avoid]

[Bug rtl-optimization/58662] [4.9 Regression] wrong code at -O2 and -O3 on x86_64-linux-gnu (in 64-bit mode)

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|ebotcazou at gcc dot gnu.org       |
           Assignee|unassigned at gcc dot gnu.org      |ebotcazou at gcc dot gnu.org

--- Comment #9 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Investigating.

[Bug rtl-optimization/58662] [4.9 Regression] wrong code at -O2 for dumb condition

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

--- Comment #10 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Author: ebotcazou
Date: Sun Oct 13 13:29:28 2013
New Revision: 203498

URL: http://gcc.gnu.org/viewcvs?rev=203498&root=gcc&view=rev
Log:
    PR rtl-optimization/58662
    * combine.c (try_combine): Take into account death nodes on I2 when
    splitting a PARALLEL of two independent SETs.  Fix dump message.

Added:
    trunk/gcc/testsuite/gcc.c-torture/execute/pr58662.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/combine.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/58662] [4.9 Regression] wrong code at -O2 for dumb condition

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58662

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #11 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Thanks for reporting the problem.