[tree-ssa] Write to global memory not hoisted out of loop

[tree-ssa] Write to global memory not hoisted out of loop

[Richard Guenther]

Hi!

Writes to global memory are not hoisted out of a computation loop. F.e.

double tmp;
void foo(double * __restrict__ x, double * __restrict__ y, double *
__restrict__ z, int n)
{
  int i;
  for (i=0; i<n; ++i) {
    tmp = y[i]*z[i];
    x[i] = tmp;
  }
}

is compiled to (-O2 -fno-trapping-math):

...
.L4:
        leal    0(,%edx,8), %eax        #, tmp66
        incl    %edx    # i
        fldl    (%eax,%ebx)     #* z
        cmpl    %ecx, %edx      # n, i
        fmull   (%eax,%esi)     #* y
        fstl    tmp     # tmp
        fstpl   (%eax,%edi)     #* x
        jl      .L4     #,
...

note that the load from tmp is optimized, but the store is still inside
the loop.  The store would be necessary in case of trapping math
operations to make side-effects visible, but I thought, -fno-trapping-math
should get rid of it.  Making tmp static doesn't help either, while here
removing the store would be valid even in case of trapping math.  Turning
on points-to doesn't help either.  Intel compiler moves the store in any
case (which seems to be a bug).

Is there already a PR about this pessimization? (Couldn't find one)

Richard.

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Steven Bosscher]

On Apr 30, 2004 01:14 PM, Richard Guenther <rguenth@tat.physik.uni-tuebingen.de> wrote:

> Writes to global memory are not hoisted out of a computation loop.

Try this with the LNO branch, iirc it can do this.


> Is there already a PR about this pessimization? (Couldn't find one)

Why is this a pessimization, does gcc 3.4.0 do what you want?  Looks
like just a missed opportunity.  I am more surprised that making tmp
static doesn't help...

Gr.
Steven

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Richard Guenther]

On Fri, 30 Apr 2004, Steven Bosscher wrote:

> On Apr 30, 2004 01:14 PM, Richard Guenther <rguenth@tat.physik.uni-tuebingen.de> wrote:
>
> > Writes to global memory are not hoisted out of a computation loop.
>
> Try this with the LNO branch, iirc it can do this.

Not as of gcc (GCC) 3.5-tree-ssa-lno 20040430 (merged 20040425) (-O2
-ftree-loop-optimize -ffast-math), neither with static nor without.

> > Is there already a PR about this pessimization? (Couldn't find one)
>
> Why is this a pessimization, does gcc 3.4.0 do what you want?  Looks
> like just a missed opportunity.  I am more surprised that making tmp
> static doesn't help...

Yes, I assumed tree-ssa can do this with some part of DOM - but
appearantly not.  No version of gcc can do this, but I think of
pessimization as of "worse than best possible", not as of a regression
towards an older release.

Richard.

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Zdenek Dvorak]

Hello,

> > > Writes to global memory are not hoisted out of a computation loop.
> >
> > Try this with the LNO branch, iirc it can do this.
> 
> Not as of gcc (GCC) 3.5-tree-ssa-lno 20040430 (merged 20040425) (-O2
> -ftree-loop-optimize -ffast-math), neither with static nor without.

tree-ssa aliasing probably does not take restricted pointers into
account, since it decides that all the accesses in the loop may alias.
Consequently none of the optimizations that could cause the store
to be hoisted out detects the possibility.

Zdenek

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Diego Novillo]

On Fri, 2004-04-30 at 16:57, Zdenek Dvorak wrote:

> tree-ssa aliasing probably does not take restricted pointers into
> account, since it decides that all the accesses in the loop may alias.
>
Correct.  We don't check restricted pointers.  This is another item in
my post merge todo.


Diego.

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Richard Kenner]

    Correct.  We don't check restricted pointers.  This is another item in
    my post merge todo.

Do you call the normal get_alias_set code?  Is TYPE_REF_CAN_ALIAS_ALL
respected?

Re: [tree-ssa] Write to global memory not hoisted out of loop

[Diego Novillo]

On Fri, 2004-04-30 at 17:17, Richard Kenner wrote:

> Do you call the normal get_alias_set code?  Is TYPE_REF_CAN_ALIAS_ALL
> respected?
>
Yes.  My mistake.  We base all the TBAA decisions on get_alias_set. 
It's in the flow-sensitive points-to code where we don't check
'restrict'.


Diego.