[llvm] [LoongArch] fix: Support LoongArch atomic memory operations with monotonic semantics (PR #112681)

[Lu Weining via llvm-commits]

SixWeining wrote:

About `atomic_load_sub`, seems we have something to optimize:

testcase:
```
#include <stdatomic.h>

int *fp2a(int **p) {
  return __atomic_fetch_sub(p, 1, memory_order_relaxed);
}
```

gcc:
```
        addi.w  $r13,$r0,-1                     # 0xffffffffffffffff
        or      $r12,$r4,$r0
        amadd.d $r4,$r13,$r12,0
```

clang:
```
        ori     $a1, $zero, 1
        sub.d   $a2, $zero, $a1
        amadd_db.d      $a1, $a2, $a0
        move    $a0, $a1
```

But this can be done in a separate PR.

https://github.com/llvm/llvm-project/pull/112681