[AArch64] Redundant `cmp` not eliminated

[Kmeakin]

https://godbolt.org/z/av7vTE7as

#include <cstdint>
#include <numeric>

typedef uint32_t u32;

u32 midpoint_u32(u32 a, u32 b) { return std::midpoint(a, b); }

Generated code:

_Z12midpoint_u32jj:                     // @_Z12midpoint_u32jj
        cmp     w0, w1
        csel    w8, w0, w1, lo
        csel    w9, w0, w1, hi
        cmp     w0, w1
        sub     w8, w9, w8
        lsr     w8, w8, #1
        cneg    w8, w8, hi
        add     w0, w8, w0
        ret

The 2nd cmp is unnecessary: the flags register still contains the result of the first cmp instruction

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

https://godbolt.org/z/av7vTE7as ```c++ #include <cstdint> #include <numeric>

typedef uint32_t u32;

u32 midpoint_u32(u32 a, u32 b) { return std::midpoint(a, b); }

Generated code:
```asm
_Z12midpoint_u32jj:                     // @<!-- -->_Z12midpoint_u32jj
        cmp     w0, w1
        csel    w8, w0, w1, lo
        csel    w9, w0, w1, hi
        cmp     w0, w1
        sub     w8, w9, w8
        lsr     w8, w8, #<!-- -->1
        cneg    w8, w8, hi
        add     w0, w8, w0
        ret

The 2nd cmp is unnecessary: the flags register still contains the result of the first cmp instruction

[davemgreen]

I would guess this is because the SUB is emitted as SUBS and the ordering of the instructions is slightly different pre-scheduling.

[ginsbach]

This seems to have been fixed by #92576. Certainly, the link

https://godbolt.org/z/av7vTE7as

#include <cstdint>
#include <numeric>

typedef uint32_t u32;

 u32 midpoint_u32(u32 a, u32 b) { return std::midpoint(a, b); }

shows the following output now:

_Z12midpoint_u32jj:
        subs    w8, w0, w1
        cneg    w8, w8, ls
        lsr     w8, w8, #1
        cneg    w8, w8, hi
        add     w0, w8, w0
        ret

I suggest we close this issue.