Optimize vadd_s64(vget_low_s64(x), vget_high_s64(x)) as vaddvq_s64(x)

[llvmbot]

Bugzilla Link	25612
Version	trunk
OS	All
Attachments	Test code with sample functions and test they are identical
Reporter	LLVM Bugzilla Contributor

Extended Description

On Aarch64, clang/LLVM misses an optimization for vadd_s64(vget_low_s64(x), vget_high_s64(x)). It can be emitted as addp.2d, in the same way as vaddvq_s64.

The reason someone would write the former rather than the latter is that the former is also valid armv7 Neon intrinsic code, whereas the latter is aarch64-only. (This arose in actual code, Neon optimizations for the Opus audio codec.)

See attached test code. The two test functions optimize as:

func1:
ext v1.16b, v0.16b, v0.16b, #​8
add d0, d0, d1
fmov x0, d0
ret

func2:
addp d0, v0.2d
fmov x0, d0
ret

even though they have identical behavior.

[c-rhodes]

reproducer: https://godbolt.org/z/5KEP1Y6av

func1:
        ext     v1.16b, v0.16b, v0.16b, #8
        add     d0, d0, d1
        fmov    x0, d0
        ret

func2:
        addp    d0, v0.2d
        fmov    x0, d0
        ret

[davemgreen]

It should maybe be being SLP vectorized into a vecreduce.add, but the vector types are probably preventing that. Adding a pattern in the backend should be simple enough.