[Bug c/94212] New: [AARCH64] [Regression] Incorrect vectorization of loop with FP calculations

[Bug c/94212] New: [AARCH64] [Regression] Incorrect vectorization of loop with FP calculations

[dpochepk at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

            Bug ID: 94212
           Summary: [AARCH64] [Regression] Incorrect vectorization of loop
                    with FP calculations
           Product: gcc
           Version: tree-ssa
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dpochepk at gmail dot com
  Target Milestone: ---

Created attachment 48054
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48054&action=edit
example application returning different result for O3 and O2

Example application (FP polynomial loop calculations) gives different result
for O2 and O3 optimizations. Different is 1 ulp, so it might be some kind of
rounding error (unsafe math leaked?).
"-O3 -fno-tree-vectorize" gives correct result.

This issue seems to affect aarch64-only (at least x86_64 is fine).
Tried several gcc versions:

trunk: affected
gcc8.3: affected
gcc7.4: not affected

(I haven't investigated assembly)

Example application is in attachment. Method "foo" has vectorized loop, which
is probably the trigger for this bug.

[Bug middle-end/94212] [AARCH64] [Regression] Incorrect vectorization of loop with FP calculations

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c                           |middle-end

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I suspect there is an Fused multiple-add being used for vectorized version
while not for the non-vectorized version.

[Bug middle-end/94212] [AARCH64] [Regression] Incorrect vectorization of loop with FP calculations

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>(at least x86_64 is fine)
No, just FMA is not enabled by default.  
If I use -march=skylake-avx512 , I get the same answers as on aarch64_64.

[Bug middle-end/94212] Incorrect vectorization of loop with FP calculations

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #2)
> >(at least x86_64 is fine)
> No, just FMA is not enabled by default.  
> If I use -march=skylake-avx512 , I get the same answers as on aarch64_64.

Note I am running on Intel(R) Xeon(R) D-2146NT CPU @ 2.30GHz.

[Bug middle-end/94212] Incorrect vectorization of loop with FP calculations

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note also using -ffp-contract=off will also change the value.

[Bug tree-optimization/94212] [8/9/10 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |7.4.0
            Summary|Incorrect vectorization of  |[8/9/10 Regression]
                   |loop with FP calculations   |Incorrect vectorization of
                   |                            |loop with FP calculations
          Component|middle-end                  |tree-optimization
            Version|tree-ssa                    |10.0
   Target Milestone|---                         |8.5
                 CC|                            |rguenth at gcc dot gnu.org,
                   |                            |rsandifo at gcc dot gnu.org
             Target|                            |aarch64

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
The vectorizer vectorizes the reduction in-order but appearantly sth goes wrong
there.

[Bug tree-optimization/94212] [8/9/10 Regression] Incorrect vectorization of loop with FP calculations

[dpochepk at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #6 from Dmitrij Pochepko <dpochepk at gmail dot com> ---
Just checked: non-vectorized assembly for aarch64 (O2) is using fmadd and fmsub
intensively.

[Bug tree-optimization/94212] [8/9/10 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Dmitrij Pochepko from comment #6)
> Just checked: non-vectorized assembly for aarch64 (O2) is using fmadd and
> fmsub intensively.

Try with -ffp-contract=off then.  Note due to effective unrolling of
the loop with vectorization we might end up forming "different" fmadd
groups.  So you might also want to check whether the vectorized loop still
sees fmadd use.

[Bug tree-optimization/94212] [8/9/10 Regression] Incorrect vectorization of loop with FP calculations

[dpochepk at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #8 from Dmitrij Pochepko <dpochepk at gmail dot com> ---
(In reply to Richard Biener from comment #7)
> (In reply to Dmitrij Pochepko from comment #6)
> > Just checked: non-vectorized assembly for aarch64 (O2) is using fmadd and
> > fmsub intensively.
> 
> Try with -ffp-contract=off then.  Note due to effective unrolling of
> the loop with vectorization we might end up forming "different" fmadd
> groups.  So you might also want to check whether the vectorized loop still
> sees fmadd use.

-O2 -ffp-contract=off
-O3 -ffp-contract=off
produce same calculation result as -O2


regarging assembly:
vectorized version is using fmla and fmls, which is vectorized version of
multiply-add/sub. It's hard to say the difference in how multiplications and
additions/subtractions are grouped without detailed step-by-step comparison
though.

[Bug tree-optimization/94212] [8/9/10/11 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
So I'm not sure this is actually a bug - the vectorizer in-order reduction
vectorization changes the FMA groups we can form (but they are not required by
the standard since they are not explicit fma () calls).

So I'd argue this report is invalid or at least I don't see a good way to fix
it
(we'd have to form FMAs before vectorization maybe).

Richard?

[Bug tree-optimization/94212] [8/9/10/11 Regression] Incorrect vectorization of loop with FP calculations

[qianjh at cn dot fujitsu.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

Qian Jianhua <qianjh at cn dot fujitsu.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |qianjh at cn dot fujitsu.com

--- Comment #10 from Qian Jianhua <qianjh at cn dot fujitsu.com> ---
I simplified the testcase as below. 
----------------------------------
#include <stdio.h>

#define POLY(arg) (1.0d - 2.0d * arg + 3.0d * arg * arg - 4.0 * arg * arg *
arg)
double v[2]={1000006.0d,1000008.0d};

int main(int argc, char** argv) {
  double res = 0.0d;
  res = v[1]*POLY(v[0]);
  printf("RESULT:%lf\n", res);
  return 0;;
}
----------------------------------

The result is also different with/without -ffp-contract=off. 
Assembly has no problem.

So I think there are differences in FP accuracy between mul+add and fmadd.
They are not equivalent for some specific numbers.

[Bug tree-optimization/94212] [8/9/10/11 Regression] Incorrect vectorization of loop with FP calculations

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

--- Comment #11 from rguenther at suse dot de <rguenther at suse dot de> ---
On Fri, 26 Mar 2021, qianjh at cn dot fujitsu.com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212
> 
> Qian Jianhua <qianjh at cn dot fujitsu.com> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>                  CC|                            |qianjh at cn dot fujitsu.com
> 
> --- Comment #10 from Qian Jianhua <qianjh at cn dot fujitsu.com> ---
> I simplified the testcase as below. 
> ----------------------------------
> #include <stdio.h>
> 
> #define POLY(arg) (1.0d - 2.0d * arg + 3.0d * arg * arg - 4.0 * arg * arg *
> arg)
> double v[2]={1000006.0d,1000008.0d};
> 
> int main(int argc, char** argv) {
>   double res = 0.0d;
>   res = v[1]*POLY(v[0]);
>   printf("RESULT:%lf\n", res);
>   return 0;;
> }
> ----------------------------------
> 
> The result is also different with/without -ffp-contract=off. 
> Assembly has no problem.
> 
> So I think there are differences in FP accuracy between mul+add and fmadd.
> They are not equivalent for some specific numbers.

Of course they are not.  fmadd elides the intermediate rounding step
of the multiplication.  The result can differ in 1ulp

[Bug tree-optimization/94212] [9/10/11/12 Regression] Incorrect vectorization of loop with FP calculations

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|8.5                         |9.4

--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 8 branch is being closed.

[Bug tree-optimization/94212] [9/10/11/12 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|9.4                         |9.5

--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 9.4 is being released, retargeting bugs to GCC 9.5.

[Bug tree-optimization/94212] [10/11/12/13 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|9.5                         |10.4

--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 9 branch is being closed

[Bug tree-optimization/94212] [10/11/12/13 Regression] Incorrect vectorization of loop with FP calculations

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|10.4                        |10.5

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 10.4 is being released, retargeting bugs to GCC 10.5.

[Bug tree-optimization/94212] [10/11/12/13 Regression] Incorrect vectorization of loop with FP calculations

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94212

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

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

--- Comment #16 from Richard Biener <rguenth at gcc dot gnu.org> ---
Invalid.