[Bug tree-optimization/95960] New: GCC should re-vectorize vector code with larger VF

[Bug tree-optimization/95960] New: GCC should re-vectorize vector code with larger VF

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95960

            Bug ID: 95960
           Summary: GCC should re-vectorize vector code with larger VF
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

For the following simple case GCC should vectorize it using AVX:

typedef double __attribute__((vector_size(16))) v2df;
double a[4];
double b[4];
void f()
{
  *(v2df *)&b[0] += *(v2df *)&a[0];
  *(v2df *)&b[2] += *(v2df *)&a[2];
}

Similar opportunities may exist for loops where also Neon -> SVE transforms
may be profitable.  Starting with BB vectorization might be easier since
you only need to consider SLP.

[Bug tree-optimization/95960] GCC should re-vectorize vector code with larger VF

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95960

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 48801
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48801&action=edit
start

So quick experiments show where we need to do adjustments.  We do have
an explicit SLP_TREE_LANES we could use for "true" scalar lanes as compared
to the number of "scalar" stmts but uses might have been transitioned not
exactly along this boundary.  Then there's of course 'group_size' which
can mean either depending on context.

The biggest churn is likely in dataref group analysis where we IMHO
also want a "lanes" representation.

[Bug tree-optimization/95960] GCC should re-vectorize vector code with larger VF

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95960

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Severity|normal                      |enhancement

[Bug tree-optimization/95960] GCC should re-vectorize vector code with larger VF

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95960

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-01-26
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

Even V2SF -> V4SF would be useful for aarch64:
```
typedef float __attribute__((vector_size(8))) v2sf;
float a[4];
float b[4];
void f()
{
  *(v2sf *)&b[0] += *(v2sf *)&a[0];
  *(v2sf *)&b[2] += *(v2sf *)&a[2];
}

```