[AArch64] Use SVE XAR for fixed-length operations.

[davemgreen]

There is a Neon SHA3 v2i64 XAR operation, but not for v4i32, v8i16 and v16i8. If sve2-sha3 is available we can use the SVE instructions instead.

https://godbolt.org/z/9hdqKoWMx (G1 and F1 are already OK).
vs with scalable vectors: https://godbolt.org/z/GhazeoaWY

https://godbolt.org/z/fejTchexj

typedef char __attribute__ ((vector_size (16))) v16qi;
typedef unsigned short __attribute__ ((vector_size (16))) v8hi;
typedef unsigned int __attribute__ ((vector_size (16))) v4si;
typedef unsigned long long __attribute__ ((vector_size (16))) v2di;
typedef char __attribute__ ((vector_size (8))) v8qi;
typedef unsigned short __attribute__ ((vector_size (8))) v4hi;
typedef unsigned int __attribute__ ((vector_size (8))) v2si;

v2di
G1 (v2di r) {
    return (r >> 39) | (r << 25);
}

v4si
G2 (v4si r) {
    return (r >> 23) | (r << 9);
}

v8hi
G3 (v8hi r) {
    return (r >> 5) | (r << 11);
}

v16qi
G4 (v16qi r)
{
  return (r << 2) | (r >> 6);
}

v2si
G5 (v2si r) {
    return (r >> 22) | (r << 10);
}

v4hi
G6 (v4hi r) {
    return (r >> 7) | (r << 9);
}

v8qi
G7 (v8qi r)
{
  return (r << 3) | (r >> 5);
}

See #137162, this is an extension to that issue.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: David Green (davemgreen)

There is a Neon SHA3 v2i64 XAR operation, but not for v4i32, v8i16 and v16i8. If sve2-sha3 is available we can use the SVE instructions instead.

https://godbolt.org/z/9hdqKoWMx (G1 and F1 are already OK).
vs with scalable vectors: https://godbolt.org/z/GhazeoaWY

https://godbolt.org/z/fejTchexj

typedef char __attribute__ ((vector_size (16))) v16qi;
typedef unsigned short __attribute__ ((vector_size (16))) v8hi;
typedef unsigned int __attribute__ ((vector_size (16))) v4si;
typedef unsigned long long __attribute__ ((vector_size (16))) v2di;
typedef char __attribute__ ((vector_size (8))) v8qi;
typedef unsigned short __attribute__ ((vector_size (8))) v4hi;
typedef unsigned int __attribute__ ((vector_size (8))) v2si;

v2di
G1 (v2di r) {
    return (r >> 39) | (r << 25);
}

v4si
G2 (v4si r) {
    return (r >> 23) | (r << 9);
}

v8hi
G3 (v8hi r) {
    return (r >> 5) | (r << 11);
}

v16qi
G4 (v16qi r)
{
  return (r << 2) | (r >> 6);
}

v2si
G5 (v2si r) {
    return (r >> 22) | (r << 10);
}

v4hi
G6 (v4hi r) {
    return (r >> 7) | (r << 9);
}

v8qi
G7 (v8qi r)
{
  return (r << 3) | (r >> 5);
}

See #137162, this is an extension to that issue.

[Rajveer100]

The third link needs all architectural features to be enabled, else we don't see the transformation for G1:

https://godbolt.org/z/hWqa5T9z3

[Rajveer100]

The first two links also need correction:

-mtriple=aarch64 -mattr=+sve2,+sve2-sha3,+sha3

=aarch64 is needed for the transformation.

https://godbolt.org/z/oGPKnYrob

https://godbolt.org/z/neszhddn5

[davemgreen]

Godbolt uses a cache and sometimes need to refresh when changes happen to trunk. There is a little refresh button on the bottom left of the compilation window (it should not be cached to the updated version).