LLVM is too quick to run `LowerExpectIntrinsicPass` requiring redundant `expect` calls

https://godbolt.org/z/dE7eGzojn

```c++
bool is_multiple_of(unsigned a, unsigned b) {
return __builtin_expect(a % b == 0, 1);
}

unsigned test(unsigned n, unsigned* out) {
auto b = is_multiple_of(n, 2);

if (b) {
return 0;
} else {
*out = 0;

return 1;
}
}

unsigned test2(unsigned n, unsigned* out) {
auto b = is_multiple_of(n, 2);

// redundant expect call
if (__builtin_expect(b, 1)) {
return 0;
} else {
*out = 0;

return 1;
}
}
```

```assembly
test(unsigned int, unsigned int*):
xor eax, eax
test dil, 1
je .LBB1_2
mov dword ptr [rsi], 0
mov eax, 1
.LBB1_2:
ret

test2(unsigned int, unsigned int*):
xor eax, eax
test dil, 1
jne .LBB2_1
ret
.LBB2_1:
mov dword ptr [rsi], 0
mov eax, 1
ret
```

Start with the Compiler Explorer reproducer at https://godbolt.org/z/dE7eGzojn and investigate the LowerExpectIntrinsicPass named in the issue. Compare the generated code for test and test2; done means the redundant expect call no longer changes the resulting control flow or code unnecessarily.

Written by the indexing model from the issue text.