https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82189
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note starting GCC 14 on aarch64 we get:
ldp s31, s30, [x1]
add x1, x2, 4
dup v0.4s, v0.s[0]
ld1 {v30.s}[1], [x1]
ld1 {v31.s}[1], [x2]
zip1 v30.4s, v31.4s, v30.4s
fdiv v0.4s, v30.4s, v0.4s
str q0, [x0]
And on the trunk we get:
ldp s31, s30, [x1]
dup v0.4s, v0.s[0]
ldr s29, [x2, 4]
ld1 {v31.s}[1], [x2]
uzp1 v30.2s, v30.2s, v29.2s
zip1 v30.4s, v31.4s, v30.4s
fdiv v0.4s, v30.4s, v0.4s
str q0, [x0]
Which is slightly worse?
This is all from:
```
_1 = *b_9(D);
_3 = MEM[(float *)b_9(D) + 4B];
_5 = *c_15(D);
_7 = MEM[(float *)c_15(D) + 4B];
_18 = {_1, _3, _5, _7};
```
```
#define vec8 __attribute__((vector_size(8)))
#define vec16 __attribute__((vector_size(16)))
vec16 float f1(float *restrict a, float * restrict b)
{
vec8 float t = {a[0], a[1]};
vec8 float t1 = {b[0], b[1]};
return __builtin_shufflevector(t, t1, 0, 1, 2, 3);
}
vec16 float f2(float *restrict a, float * restrict b)
{
vec16 float t = {a[0], a[1], b[0], b[1]};
return t;
}
```
We can optimize f1 but not f2.
