[Bug c/52407] New: sse2 simd uint32_t and int64_t and stack variable initialization

[Bug c/52407] New: sse2 simd uint32_t and int64_t and stack variable initialization

[pcpa at mandriva dot com.br]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

             Bug #: 52407
           Summary: sse2 simd uint32_t and int64_t and stack variable
                    initialization
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: pcpa@mandriva.com.br


Created attachment 26766
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26766
simple test

The attached test case works if adapted for smaller integer
types (initializing more fields through the pointer), and also
works if declaring either "mul_vl_l" variables "p" or "w" static
or global.

  If compiled with -O0 it also works, but -O1 or greater
generates incorrect result. If not casting and initializing
through the pointer, as in the "#if 0" code it also works.

expected output:
$ ./a.out
2 4 6 8 10 12 14 16 

generated output:
$ ./a.out
0 4 0 8 0 12 0 16

[Bug target/52407] sse2 simd uint32_t and int64_t and stack variable initialization

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

--- Comment #1 from Uros Bizjak <ubizjak at gmail dot com> 2012-02-27 23:34:33 UTC ---
This looks like an aliasing violation to me:

    vl_t         w;
    int64_t     *p = (int64_t *)&w;
    p[0] = p[1] = x;

Using following code, the test works OK:

  union {
    vl_t w;
    int64_t p[2];
  } z;

  z.p[0] = z.p[1] = x;

  while (m--)
    *u++ = *v++ * z.w;

[Bug target/52407] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-02-28
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 10:39:53 UTC ---
It looks like a bug in the way we expand the vector init:

;; MEM[(int64_t *)&w + 8B] = 2;

(insn 39 38 40 (set (reg:DI 102)
        (const_int 2 [0x2])) t.c:19 -1
     (nil))

(insn 40 39 41 (set (reg:DI 103)
        (vec_select:DI (reg/v:V2DI 98 [ w ])
            (parallel [
                    (const_int 0 [0])
                ]))) t.c:19 -1
     (nil))

(insn 41 40 0 (set (reg/v:V2DI 98 [ w ])
        (vec_concat:V2DI (reg:DI 102)
            (reg:DI 103))) t.c:19 -1
     (nil))

^^^ should be vec_concat 103 102

;; MEM[(int64_t *)&w] = 2;

(insn 42 41 43 (set (reg:DI 104)
        (const_int 2 [0x2])) t.c:19 -1
     (nil))

(insn 43 42 44 (set (reg:DI 105)
        (vec_select:DI (reg/v:V2DI 98 [ w ])
            (parallel [
                    (const_int 1 [0x1])
                ]))) t.c:19 -1
     (nil))

(insn 44 43 0 (set (reg/v:V2DI 98 [ w ])
        (vec_concat:V2DI (reg:DI 105)
            (reg:DI 104))) t.c:19 -1
     (nil))

^^^  should be vec_concat 104 105

no?

[Bug target/52407] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 11:32:34 UTC ---
Hmm,

ix86_expand_vector_set

has for the V2DI case

      ix86_expand_vector_extract (false, tmp, target, 1 - elt);
      if (elt == 0)
        tmp = gen_rtx_VEC_CONCAT (mode, tmp, val);
      else
        tmp = gen_rtx_VEC_CONCAT (mode, val, tmp);

vs. for the V2DF case:

        tmp = gen_rtx_VEC_SELECT (inner_mode, target, tmp);

        if (elt == 0)
          op0 = val, op1 = tmp;
        else
          op0 = tmp, op1 = val;

        tmp = gen_rtx_VEC_CONCAT (mode, op0, op1);

looks the V2DI case has swapped operands.

[Bug target/52407] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

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

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 11:34:36 UTC ---
I'm testing a patch.  The V2DF/SImode case is also bogus.

[Bug target/52407] [4.7 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.6.2
   Target Milestone|---                         |4.7.0
            Summary|sse2 simd uint32_t and      |[4.7 Regression] sse2 simd
                   |int64_t and stack variable  |uint32_t and int64_t and
                   |initialization              |stack variable
                   |                            |initialization

[Bug target/52407] [4.7 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-02-28 12:02:18 UTC ---
Triggered by http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=184435
Testcase:
/* PR target/52407 */

extern void abort (void);

typedef long long V __attribute__ ((vector_size (16)));
V ul[4], vl[4] = { {1, 2}, {3, 4}, {5, 6}, {7, 8} };

static void
foo (V *u, V *v, long long x, int m)
{
  V w;
  long long *p = (long long *) &w;
  p[0] = p[1] = x;
  while (m--)
    *u++ = *v++ * w;
}

int
main ()
{
  int i;
  long long *pl;

  pl = (long long *) &ul;
  foo (ul, vl, 2, 4);
  for (i = 0; i < 8; i++)
    if (pl[i] != 2 * (i + 1))
      abort ();
  return 0;
}

[Bug target/52407] [4.7 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 15:28:39 UTC ---
Author: rguenth
Date: Tue Feb 28 15:28:32 2012
New Revision: 184627

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=184627
Log:
2012-02-28  Richard Guenther  <rguenther@suse.de>

    PR target/52407
    * config/i386/i386.c (ix86_expand_vector_set): Fix element
    ordering for the VEC_CONCAT for two element vectors for
    V2SFmode, V2SImode and V2DImode.

    * gcc.dg/torture/pr52407.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/torture/pr52407.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/testsuite/ChangeLog

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.7.0
   Target Milestone|4.7.0                       |4.6.4
            Summary|[4.7 Regression] sse2 simd  |[4.6 Regression] sse2 simd
                   |uint32_t and int64_t and    |uint32_t and int64_t and
                   |stack variable              |stack variable
                   |initialization              |initialization

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 15:29:53 UTC ---
Fixed on trunk.  I'll backport it for 4.6.4 as it is latent on the branches.

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

Martin Jambor <jamborm at gcc dot gnu.org> changed:

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

--- Comment #8 from Martin Jambor <jamborm at gcc dot gnu.org> 2012-02-28 18:09:48 UTC ---
*** Bug 52419 has been marked as a duplicate of this bug. ***

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[pcpa at mandriva dot com.br]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

Paulo César Pereira de Andrade <pcpa at mandriva dot com.br> changed:

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

--- Comment #9 from Paulo César Pereira de Andrade <pcpa at mandriva dot com.br> 2012-02-29 17:54:34 UTC ---
(In reply to comment #7)
> Fixed on trunk.  I'll backport it for 4.6.4 as it is latent on the branches.

Thanks. I reported the problem because I thought it could
hit somewhere else and I have some code that way, that was
working since gcc 4.4, but the proper syntax to initialize
the local variable should be, in the sample test case:
    vl_t     w = {x,x};
what should make it easier for the compiler to understand
what is being done.

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

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

--- Comment #10 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-03-01 09:53:46 UTC ---
Re-open so I don't forget about the backporting.

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

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

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

--- Comment #12 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-05-22 09:20:41 UTC ---
Fixed.

[Bug target/52407] [4.6 Regression] sse2 simd uint32_t and int64_t and stack variable initialization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52407

--- Comment #11 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-05-22 09:20:24 UTC ---
Author: rguenth
Date: Tue May 22 09:20:15 2012
New Revision: 187763

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=187763
Log:
2012-05-22  Richard Guenther  <rguenther@suse.de>

        Backport from mainline
        2012-02-28  Richard Guenther  <rguenther@suse.de>

    PR target/52407
    * config/i386/i386.c (ix86_expand_vector_set): Fix element
    ordering for the VEC_CONCAT for two element vectors for
    V2SFmode, V2SImode and V2DImode.

    * gcc.dg/torture/pr52407.c: New testcase.

Added:
    branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/torture/pr52407.c
Modified:
    branches/gcc-4_6-branch/gcc/ChangeLog
    branches/gcc-4_6-branch/gcc/config/i386/i386.c
    branches/gcc-4_6-branch/gcc/testsuite/ChangeLog