[Bug rtl-optimization/65235] New: [4.8, 4.9, 5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[Bug rtl-optimization/65235] New: [4.8, 4.9, 5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

            Bug ID: 65235
           Summary: [4.8, 4.9, 5 Regression] Simplifying vec_select of
                    vec_concat miscompiles when first element of
                    vec_concat is const_int
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
            Target: aarch64

This aarch64 intrinsics testcase aborts.

#include "arm_neon.h"

int main (int argc, char** argv)
{
  int64x1_t val1;
  int64x1_t val2;
  int64x1_t val3;
  uint64x1_t val13;
  uint64x2_t val14;
  uint64_t got;
  uint64_t exp;
  val1 = vcreate_s64(UINT64_C(0xffffffff80008000));
  val2 = vcreate_s64(UINT64_C(0x0000f38d00000000));
  val3 = vcreate_s64(UINT64_C(0xffff7fff0000809b));
  //  Expect: "val13" = 8000000000001553
  val13 = vcreate_u64 (UINT64_C(0x8000000000001553));
  //  Expect: "val14" = 0010 0000 0000 0002 0000 0000 0000 0000
  val14 = vcombine_u64(vcgt_s64(vqrshl_s64(val1, val2),
                vshr_n_s64(val3, 18)),
               vshr_n_u64(val13, 11));
  /* Should be 0000000000000000.  */
  got = vgetq_lane_u64(val14, 0);
  exp = 0;
  if(exp != got)
    __builtin_abort ();
}



Investigation shows that the problem is in the simplify-rtx machinery:
Combine tries to combine:
 (insn 72 71 73 2 (set (reg:V2DI 117 [ D.18177 ])
         (vec_concat:V2DI (reg:DI 176 [ D.18179 ])
           (reg:DI 114 [ D.18168 ]))) 
      (expr_list:REG_DEAD (reg:DI 176 [ D.18179 ])
      (expr_list:REG_DEAD (reg:DI 114 [ D.18168 ])

and

 (insn 104 102 105 2 (set (reg:DI 193 [ D.18168 ])
         (vec_select:DI (reg:V2DI 117 [ D.18177 ])
             (parallel [
                     (const_int 0 [0])
                 ])))
      (expr_list:REG_DEAD (reg:V2DI 117 [ D.18177 ])
         (nil)))

but ends up generating:
(set (reg:DI 193 [ D.18168 ])
    (reg:DI 114 [ D.18168 ]))

i.e. it picks element 1 instead of the requested element 0.

This happens during combine where it tries to to simplify a vec_select 0 of the
intermediate rtx:
(vec_concat:V2DI (const_int -1 [0xffffffffffffffff])
    (reg:DI 114 [ D.18166 ]))


The relevant code in simplify-rtx has been there for some time and is broken
for all release branches. I have a patch in testing...

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |5.0
            Summary|[4.8, 4.9, 5 Regression]    |[4.8/4.9/5 Regression]
                   |Simplifying vec_select of   |Simplifying vec_select of
                   |vec_concat miscompiles when |vec_concat miscompiles when
                   |first element of vec_concat |first element of vec_concat
                   |is const_int                |is const_int

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|5.0                         |4.8.5

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #2 from ktkachov at gcc dot gnu.org ---
(In reply to Aldy Hernandez from comment #1)
> Hi.  The known to work field is empty.  Is this a regression?
> 
> I can't reproduce on a cross build to --target=aarch64-linux by running cc1.
> What options causes this to fail?  Could you post a preprocessed file for
> the test?

-O2 does it for me. I didn't try anything earlier than 4.8 (since they're not
maintained) but the offending code has been there since 2005:
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3551)          rtx vec =
trueop0;
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3552)          int offset =
INTVAL (XVECEXP (trueop1, 0, 0)) * GET_MODE_SIZE (mode);
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3553) 
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3554)          /* Try to
find the element in the VEC_CONCAT.  */
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3555)          while
(GET_MODE (vec) != mode
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3556)                 &&
GET_CODE (vec) == VEC_CONCAT)
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3557)            {
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3558)             
HOST_WIDE_INT vec_size = GET_MODE_SIZE (GET_MODE (XEXP (vec, 0)));
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3559)              if
(offset < vec_size)
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3560)                vec =
XEXP (vec, 0);
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3561)              else
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3562)                {
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3563)                 
offset -= vec_size;
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3564)                  vec =
XEXP (vec, 1);
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3565)                }
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3566)              vec =
avoid_constant_pool_reference (vec);
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3567)            }
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3568) 
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3569)          if (GET_MODE
(vec) == mode)
6516cbaf (bonzini        2005-12-16 09:24:19 +0000 3570)            return vec;


and aarch64 wasn't there before 4.8...

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #3 from ktkachov at gcc dot gnu.org ---
gah, sorry, the field is not wide enough to show the output of git blame
properly...

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #5 from ktkachov at gcc dot gnu.org ---
Created attachment 34891
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34891&action=edit
Self-contained testcase

Here's a testcase with the relevant parts of arm_neon.h extracted.

This aborts for me on aarch64-none-linux-gnu and aarch64-none-elf

[Bug rtl-optimization/65235] [4.8/4.9/5 Regression] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #6 from ktkachov at gcc dot gnu.org ---

> Hmmm... fine line, but maybe this isn't a regression?

I guess it isn't a regression in the release branches, but it is wrong-code

[Bug rtl-optimization/65235] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[jgreenhalgh at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

James Greenhalgh <jgreenhalgh at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|aarch64                     |aarch64*-*-*
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-03-02
                 CC|                            |jgreenhalgh at gcc dot gnu.org
               Host|                            |*-*-*
     Ever confirmed|0                           |1

--- Comment #7 from James Greenhalgh <jgreenhalgh at gcc dot gnu.org> ---
The tescase from comment 5 fails for me at -O3 on an aarch64-none-elf cross
compiler (r221086), this also fails for native aarch64-none-linux-gnu.

As with Kyrill, I see:

(insn 16 15 17 2 (set (reg:V2DI 80 [ D.2809 ])
        (vec_concat:V2DI (reg:DI 95 [ D.2812 ])
            (reg:DI 77 [ D.2806 ]))) reduced.c:19 1386
{aarch64_combine_internaldi}
     (expr_list:REG_DEAD (reg:DI 95 [ D.2812 ])
        (expr_list:REG_DEAD (reg:DI 77 [ D.2806 ])
            (expr_list:REG_EQUAL (vec_concat:V2DI (reg:DI 95 [ D.2812 ])
                    (const_int 4503599627370498 [0x10000000000002]))
                (nil)))))

(insn 17 16 18 2 (set (reg:DI 96 [ D.2806 ])
        (vec_select:DI (reg:V2DI 80 [ D.2809 ])
            (parallel [
                    (const_int 0 [0])
                ]))) reduced.c:14 1372 {aarch64_get_lanev2di}
     (expr_list:REG_DEAD (reg:V2DI 80 [ D.2809 ])
        (nil)))
----
Combine:

Trying 16 -> 17:
Successfully matched this instruction:
(set (reg:DI 96 [ D.2806 ])
    (reg:DI 77 [ D.2806 ]))
---

Which is bogus, the correct rewrite would be to:

(set (reg:DI 96 [ D.2806 ])
    (reg:DI 95 [ D.2812 ]))

[Bug rtl-optimization/65235] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #8 from ktkachov at gcc dot gnu.org ---
Author: ktkachov
Date: Thu Mar 12 13:40:50 2015
New Revision: 221387

URL: https://gcc.gnu.org/viewcvs?rev=221387&root=gcc&view=rev
Log:
[simplify-rtx] PR 65235: Calculate element size correctly when simplifying
(vec_select (vec_concat (const_int) (...)) [...])

    PR rtl-optimization 65235
    * simplify-rtx.c (simplify_binary_operation_1, VEC_SELECT case):
    When first element of vec_concat is const_int, calculate its size
    using second element.

    PR rtl-optimization 65235
    * gcc.target/aarch64/pr65235_1.c: New test.

Added:
    trunk/gcc/testsuite/gcc.target/aarch64/pr65235_1.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/simplify-rtx.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/65235] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

ktkachov at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |5.0
      Known to fail|5.0                         |

--- Comment #9 from ktkachov at gcc dot gnu.org ---
Fixed on trunk. Will backport to branches later.

[Bug rtl-optimization/65235] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

--- Comment #10 from ktkachov at gcc dot gnu.org ---
Author: ktkachov
Date: Thu Mar 19 09:58:42 2015
New Revision: 221511

URL: https://gcc.gnu.org/viewcvs?rev=221511&root=gcc&view=rev
Log:
[simplify-rtx] PR 65235: Calculate element size correctly when simplifying
(vec_select (vec_concat (const_int) (...)) [...])

    Backport from mainline
    2015-03-12  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

    PR rtl-optimization/65235
    * simplify-rtx.c (simplify_binary_operation_1, VEC_SELECT case):
    When first element of vec_concat is const_int, calculate its size
    using second element.

    PR rtl-optimization/65235
    * gcc.target/aarch64/pr65235_1.c: New test.

Added:
    branches/gcc-4_9-branch/gcc/testsuite/gcc.target/aarch64/pr65235_1.c
Modified:
    branches/gcc-4_9-branch/gcc/ChangeLog
    branches/gcc-4_9-branch/gcc/simplify-rtx.c
    branches/gcc-4_9-branch/gcc/testsuite/ChangeLog

[Bug rtl-optimization/65235] Simplifying vec_select of vec_concat miscompiles when first element of vec_concat is const_int

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65235

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
      Known to work|                            |4.9.3
         Resolution|---                         |FIXED
   Target Milestone|4.8.5                       |4.9.3
      Known to fail|4.8.4, 4.9.3                |4.8.5, 4.9.2

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed for 4.9.3.