[Bug target/113537] New: ext should be used more for __builtin_shufflevector

[Bug target/113537] New: ext should be used more for __builtin_shufflevector

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113537

            Bug ID: 113537
           Summary: ext should be used more for __builtin_shufflevector
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64

Take:
```
#define vector4 __attribute__((vector_size(4)))
#define vector8 __attribute__((vector_size(8)))
#define vector16 __attribute__((vector_size(16)))


vector8 char f3(vector16 char a)
{
  return __builtin_shufflevector  (a, a, 1, 2, 3, 4, 5, 6, 7, 8);
}

vector8 char f2(vector16 char a)
{
  return __builtin_shufflevector  (a, a, 1, 2, 3, 4, 5, 6, 7, 0);
}
```

Currently GCC produces:
```
f3:
        adrp    x0, .LC0
        ldr     q31, [x0, #:lo12:.LC0]
        tbl     v0.16b, {v0.16b}, v31.16b
        ret
f2:
        adrp    x0, .LC1
        ldr     q31, [x0, #:lo12:.LC1]
        tbl     v0.16b, {v0.16b}, v31.16b
        ret
```

But these should be optimized to just:
```
f3:
        ext     v0.16b, v0.16b, v0.16b, #1
        ret
f2:
        ext     v0.8b, v0.8b, v0.8b, #1
        ret
```

[Bug target/113537] ext should be used more for __builtin_shufflevector

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113537

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
One gimple level thing I noticed is we produce:
```
  _1 = VEC_PERM_EXPR <a_2(D), a_2(D), { 1, 2, 3, 4, 5, 6, 7, 8, 8, 9, 10, 11,
12, 13, 14, 15 }>;
  _3 = BIT_FIELD_REF <_1, 64, 0>;
```

But Maybe that should just be:
```
_3 = VEC_PERM_EXPR <a_2(D), a_2(D), { 1, 2, 3, 4, 5, 6, 7, 8} >;
```

Which is what the vector might produce but maybe it is because
aarch64_vectorize_vec_perm_const does not know how to handle that ...

[Bug target/113537] ext should be used more for __builtin_shufflevector

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113537

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
aarch64_vectorize_vec_perm_const does not handle at all the case where the 2
modes are different.

[Bug target/113537] ext should be used more for __builtin_shufflevector

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113537

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2024-01-22

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I am going to implement at least the aarch64_vectorize_vec_perm_const  part.
I will check the other side later on.

Note this will be part of the patch set am doing for PR 113458 .

[Bug target/113537] ext should be used more for __builtin_shufflevector

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113537

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=101846

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Actually part of this is the "don't care" part, which is referenced in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101846#c10 (and maybe others).
That is something I can't fix ...