[AArch64] Suboptimal expansion of __builtin_bswap16

[Kmeakin]

https://godbolt.org/z/jcxb4n97o

uint16_t rev16(uint16_t x) { return __builtin_bswap16(x); }

expands to

rev16:
        rev     w8, w0
        lsr     w0, w8, #16
        ret

instead of

rev16:
        rev16   w0, w0
        ret

__builtin_bswap32 and __builtin_bswap64 don't have the same issue

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

https://godbolt.org/z/jcxb4n97o

uint16_t rev16(uint16_t x) { return __builtin_bswap16(x); }

expands to

rev16:
        rev     w8, w0
        lsr     w0, w8, #<!-- -->16
        ret

instead of

rev16:
        rev16   w0, w0
        ret

__builtin_bswap32 and __builtin_bswap64 don't have the same issue

[anematode]

Latest clang seems to emit rev16 here. Maybe fixed by #80036 ?

[Kmeakin]

Latest clang seems to emit rev16 here. Maybe fixed by #80036 ?

It is still suboptimal on Compiler Explorer

[landfillbaby]

this also applies to the __revsh and __rev16 ACLE intrinsics! they should generate the instruction they represent!

[adprasad-nvidia]

I've been working on a patch for this at #105375 .