Re: vectorization, scheduling and aliasing

[Richard Guenther <richard.guenther@gmail.com>]

On Mon, Apr 26, 2010 at 9:43 AM, roy rosen <roy.1rosen@gmail.com> wrote:
> Hi Richard,
>
> 2010/4/23, Richard Guenther <richard.guenther@gmail.com>:
>> On Thu, Apr 22, 2010 at 6:04 PM, roy rosen <roy.1rosen@gmail.com> wrote:
>> > Hi Richard,
>> >
>> > 2010/4/14, Richard Guenther <richard.guenther@gmail.com>:
>> >> On Wed, Apr 14, 2010 at 8:48 AM, roy rosen <roy.1rosen@gmail.com> wrote:
>> >> > Hi All,
>> >> >
>> >> > I have implemented some vectorization features in my gcc port.
>> >> >
>> >> > In the generated code for this function I can see a scheduling problem:
>> >> >
>> >> > int xxx(int* __restrict__ a, int* __restrict__ b)
>> >> > {
>> >> >    int __restrict__ i;
>> >> >    for (i = 0; i < 8; i++)
>> >> >    {
>> >> >        a[i] = b[i];
>> >> >    }
>> >> >    return 0;
>> >> > }
>> >> >
>> >> > from asmcons:
>> >> >
>> >> > (insn 17 14 18 3 a.c:15 (set (reg:V4HI 105)
>> >> >        (mem:V4HI (reg/v/f:SI 100 [ b ]) [2 S8 A64])) 115 {*movv4hi_load} (nil))
>> >> >
>> >> > (insn 18 17 19 3 a.c:15 (set (mem:V4HI (reg/v/f:SI 99 [ a ]) [2 S8 A64])
>> >> >        (reg:V4HI 105)) 116 {*movv4hi_store} (expr_list:REG_DEAD (reg:V4HI 105)
>> >> >        (nil)))
>> >> >
>> >> > (insn 19 18 20 3 a.c:15 (set (reg:V4HI 106)
>> >> >        (mem:V4HI (plus:SI (reg/v/f:SI 100 [ b ])
>> >> >                (const_int 8 [0x8])) [2 S8 A64])) 115 {*movv4hi_load}
>> >> > (expr_list:REG_DEAD (reg/v/f:SI 100 [ b ])
>> >> >        (nil)))
>> >> >
>> >> > (insn 20 19 58 3 a.c:15 (set (mem:V4HI (plus:SI (reg/v/f:SI 99 [ a ])
>> >> >                (const_int 8 [0x8])) [2 S8 A64])
>> >> >        (reg:V4HI 106)) 116 {*movv4hi_store} (expr_list:REG_DEAD (reg:V4HI 106)
>> >> >        (expr_list:REG_DEAD (reg/v/f:SI 99 [ a ])
>> >> >            (nil))))
>> >> >
>> >> >  from sched1 - dependecies list:
>> >> >   --- Region Dependences --- b 3 bb 0
>> >> > ;;      insn  code    bb   dep  prio  cost   reservation
>> >> > ;;      ----  ----    --   ---  ----  ----   -----------
>> >> > ;;       17   115     3     0     4     1   (2_slots+lsu)       : 58 20 18
>> >> > ;;       18   116     3     1     3     1   (2_slots+lsu)       : 58 19
>> >> > ;;       19   115     3     1     2     1   (2_slots+lsu)       : 58 20
>> >> > ;;       20   116     3     2     1     1   (2_slots+lsu)       : 58
>> >> > ;;       58   101     3     4     1     1   (3_slots+pcu)       :
>> >> >
>> >> > As you can see, insn 19 is dependant on 18.
>> >> > I think that this should not happen since the arrays has __restrict__.
>> >> > It might have something to do with aliasing.
>> >> >
>> >> > Are you aware of any kind of problems regarding aliasing and vectors?
>> >> > Can you think of anything else that might be wrong in my code that
>> >> > leads gcc to think that there is a dependency between insn 18 and 19?
>> >>
>> >> It completely depends on the GCC version you are using.  The
>> >> MEM_EXPRs suggest that it is maybe 4.4 or 4.5 based?  Because
>> >> they are not existant.  Try recent trunk which should work fine.
>> >>
>> >> Richard.
>> >>
>> >> > Thanks, Roy.
>> >> >
>> >>
>> >
>> > I have taken 4.5.0 and I still get the same problem.
>> > Do you have any ideas? Shouldn't the restrict tell the compiler that
>> > there is no dependency between these insns?
>>
>> It doesn't work for 4.5.0 due to PR42617.
>>
>> Richard.
>>
>> > Thanks, Roy.
>> >
>>
>
> I took snapshot 4.6-20100424 (is this version ok?) and the problem
> still occurs when I am looking at the vectorized version of the code.
> It seems ok in the nonvectorized version.
>
> What might be the problem?

Look at the -fdump-tree-optimized-alias dump, likely the defs
of the pointer SSA names that are used in the loads/stores
have their alias information dropped.  Or the vectorizer emits
ALIGN_INDIRECT_REFs on your target which we still drop.

Richard.

> Thanks, Roy.
>