[Bug tree-optimization/49006] New: [4.6/4.7 Regression] Missed vectorization due to revision 167531

[Bug tree-optimization/49006] New: [4.6/4.7 Regression] Missed vectorization due to revision 167531

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

           Summary: [4.6/4.7 Regression] Missed vectorization due to
                    revision 167531
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: dominiq@lps.ens.fr
                CC: rguenth@gcc.gnu.org


Created attachment 24250
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24250
source code

On x86_64-apple-darwin10 the attached code (an avatar of the polyhedron test
induct.f90) is no longer properly vectorized after revision 167531:

[macbook] lin/test% /opt/gcc/gcc4.6p-167530/bin/gfortran -Ofast
-ftree-vectorizer-verbose=2 induct_v4.f90
...
induct_v4.f90:1757: note: not vectorized: can't create epilog loop 1.
induct_v4.f90:1766: note: LOOP VECTORIZED.
induct_v4.f90:1711: note: LOOP VECTORIZED.
induct_v4.f90:1633: note: LOOP VECTORIZED.
induct_v4.f90:1449: note: vectorized 3 loops in function.

induct_v4.f90:2168: note: not vectorized: can't create epilog loop 1.
induct_v4.f90:2177: note: LOOP VECTORIZED.
induct_v4.f90:2095: note: LOOP VECTORIZED.
induct_v4.f90:2018: note: LOOP VECTORIZED.
induct_v4.f90:1832: note: vectorized 3 loops in function.
...
[macbook] lin/test% time a.out > /dev/null
12.677u 0.027s 0:12.70 99.9%    0+0k 0+1io 0pf+0w
[macbook] lin/test% /opt/gcc/gcc4.6p-167531/bin/gfortran -Ofast
-ftree-vectorizer-verbose=2 induct_v4.f90
...
induct_v4.f90:1728: note: not vectorized: unsupported use in stmt.
induct_v4.f90:1757: note: not vectorized: unsupported use in stmt.
induct_v4.f90:1711: note: LOOP VECTORIZED.
induct_v4.f90:1633: note: LOOP VECTORIZED.
induct_v4.f90:1449: note: vectorized 2 loops in function.

induct_v4.f90:2168: note: not vectorized: unsupported use in stmt.
induct_v4.f90:2095: note: LOOP VECTORIZED.
induct_v4.f90:2018: note: LOOP VECTORIZED.
induct_v4.f90:1832: note: vectorized 2 loops in function.
...
[macbook] lin/test% time a.out > /dev/null
21.831u 0.031s 0:21.86 100.0%    0+0k 0+0io 0pf+0w

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #1 from rguenther at suse dot de <rguenther at suse dot de> 2011-05-16 08:57:46 UTC ---
On Sun, 15 May 2011, dominiq at lps dot ens.fr wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006
> 
>            Summary: [4.6/4.7 Regression] Missed vectorization due to
>                     revision 167531
>            Product: gcc
>            Version: 4.6.0
>             Status: UNCONFIRMED
>           Severity: normal
>           Priority: P3
>          Component: tree-optimization
>         AssignedTo: unassigned@gcc.gnu.org
>         ReportedBy: dominiq@lps.ens.fr
>                 CC: rguenth@gcc.gnu.org
> 
> 
> Created attachment 24250
>   --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24250
> source code
> 
> On x86_64-apple-darwin10 the attached code (an avatar of the polyhedron test
> induct.f90) is no longer properly vectorized after revision 167531:

The patch changes inliner heuristics only.

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #2 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-05-16 09:24:28 UTC ---
> The patch changes inliner heuristics only.

Yes, but the vectorization of induct.f90 is very sensitive to inlining: see
pr34265. Did you checked the vectorization of the test case on your favorite
platform?

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> 2011-05-16 10:41:58 UTC ---
On Mon, 16 May 2011, dominiq at lps dot ens.fr wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006
> 
> --- Comment #2 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-05-16 09:24:28 UTC ---
> > The patch changes inliner heuristics only.
> 
> Yes, but the vectorization of induct.f90 is very sensitive to inlining: see
> pr34265. Did you checked the vectorization of the test case on your favorite
> platform?

I did not check it at all.  The patch causes more inlining, making the
cost of x*x the same with/without -ffast-math.

Richard.

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.6.1

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[irar at il dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

Ira Rosen <irar at il dot ibm.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |irar at il dot ibm.com

--- Comment #4 from Ira Rosen <irar at il dot ibm.com> 2011-05-19 12:33:49 UTC ---
(In reply to comment #0)

The problem is you don't see:
> induct_v4.f90:1766: note: LOOP VECTORIZED.
> induct_v4.f90:2177: note: LOOP VECTORIZED.
anymore, right?

With -Ofast there are no such loops, 1766 and 2177, they are probably being
unrolled.

I see these loops with -O2 -ftree-vectorize, and they are indeed not
vectorized, but they are not vectorized by previous version as well, AFAIU.

Ira

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #5 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-05-22 12:22:48 UTC ---
> (In reply to comment #4)
>
> The problem is you don't see:
> > induct_v4.f90:1766: note: LOOP VECTORIZED.
> > induct_v4.f90:2177: note: LOOP VECTORIZED.
> anymore, right?
>
> With -Ofast there are no such loops, 1766 and 2177, they are probably being
> unrolled.
>
> I see these loops with -O2 -ftree-vectorize, and they are indeed not
> vectorized, but they are not vectorized by previous version as well, AFAIU.
>
> Ira

I have posted at pr34265 a reduced test allowing for twelve variants of the
inner loops in induct.f90.
The vectorization is indeed depending of what the pass_complete_unrolli is
doing. In a perfect world I would expect all these variants be equally
vectorized. I have no idea if this is doable, but I think the minimal quality
of implementation would require that the variants that are not vectorized (with
a very significant impact on the run time) do not depend on the revision and be
predictable along a line "do this, don't do that".

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.6.1                       |4.6.2

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-06-27 12:32:36 UTC ---
GCC 4.6.1 is being released.

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Priority|P3                          |P2

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[irar at il dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #7 from Ira Rosen <irar at il dot ibm.com> 2011-08-01 14:47:08 UTC ---
I think that this is not really a missed-optimization. It's just that some
loops disappear because of the unrolling.

Ira

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #8 from rguenther at suse dot de <rguenther at suse dot de> 2011-08-02 08:37:28 UTC ---
On Mon, 1 Aug 2011, irar at il dot ibm.com wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006
> 
> --- Comment #7 from Ira Rosen <irar at il dot ibm.com> 2011-08-01 14:47:08 UTC ---
> I think that this is not really a missed-optimization. It's just that some
> loops disappear because of the unrolling.

Ah, ok - so SLP vectorization won't print a "vectorized loop X" if it
would vectorize the unrolled variant?  (not sure if it easily could)

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[irar at il dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #9 from Ira Rosen <irar at il dot ibm.com> 2011-08-02 09:21:26 UTC ---
Basic block SLP only prints "basic block vectorized using SLP". But I thought
we are talking about loop vectorization here, since the early unrolling is
applied only if there is an outer loop as well (correct?), so it should be the
same "LOOP VECTORIZED". And I see with on trunk:

induct.f90:1757: note: LOOP VECTORIZED.
induct.f90:2168: note: LOOP VECTORIZED.

and 1757 is the outer loop of 1766, and 2168 is the outer loop of 2177.

These loops are not vectorized with 4.6 though.

Ira

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> 2011-08-02 09:24:19 UTC ---
On Tue, 2 Aug 2011, irar at il dot ibm.com wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006
> 
> --- Comment #9 from Ira Rosen <irar at il dot ibm.com> 2011-08-02 09:21:26 UTC ---
> Basic block SLP only prints "basic block vectorized using SLP". But I thought
> we are talking about loop vectorization here, since the early unrolling is
> applied only if there is an outer loop as well (correct?), so it should be the

Yes.

[Bug tree-optimization/49006] [4.6/4.7 Regression] Missed vectorization due to revision 167531

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49006

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

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

--- Comment #11 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-09-16 15:42:03 UTC ---
This pr (as well as pr34265) seems to have been fixed between revisions 176696
and 177649:

[macbook] lin/test% /opt/gcc/gcc4.7p-176696p1/bin/gfortran -O3 -ffast-math
induct_v4.f90 -ftree-vectorizer-verbose=2
...
induct_v4.f90:1728: note: not vectorized: unsupported use in stmt.
induct_v4.f90:1711: note: LOOP VECTORIZED.
induct_v4.f90:1633: note: LOOP VECTORIZED.
induct_v4.f90:1544: note: created 2 versioning for alias checks.

induct_v4.f90:1544: note: LOOP VECTORIZED.
induct_v4.f90:1449: note: vectorized 3 loops in function.

induct_v4.f90:2095: note: LOOP VECTORIZED.
induct_v4.f90:2018: note: LOOP VECTORIZED.
induct_v4.f90:1925: note: created 2 versioning for alias checks.

induct_v4.f90:1925: note: LOOP VECTORIZED.
induct_v4.f90:1832: note: vectorized 3 loops in function.
...
[macbook] lin/test% time a.out > /dev/null
22.186u 0.023s 0:22.21 99.9%    0+0k 0+3io 0pf+0w

[macbook] lin/test% /opt/gcc/gcc4.7p-177649p2/bin/gfortran -O3 -ffast-math
induct_v4.f90 -ftree-vectorizer-verbose=2
...
induct_v4.f90:1728: note: not vectorized: unsupported data-type complex(kind=8)
induct_v4.f90:1757: note: LOOP VECTORIZED.
induct_v4.f90:1711: note: LOOP VECTORIZED.
induct_v4.f90:1633: note: LOOP VECTORIZED.
induct_v4.f90:1544: note: created 2 versioning for alias checks.

induct_v4.f90:1544: note: LOOP VECTORIZED.
induct_v4.f90:1449: note: vectorized 4 loops in function.

induct_v4.f90:2168: note: LOOP VECTORIZED.
induct_v4.f90:2095: note: LOOP VECTORIZED.
induct_v4.f90:2018: note: LOOP VECTORIZED.
induct_v4.f90:1925: note: created 2 versioning for alias checks.

induct_v4.f90:1925: note: LOOP VECTORIZED.
induct_v4.f90:1832: note: vectorized 4 loops in function.
...
[macbook] lin/test% time a.out > /dev/null
12.657u 0.022s 0:12.68 99.9%    0+0k 0+0io 0pf+0w

I am closing this pr as fixed and I'll use pr34265 to track the remaining
issues.