[GlobalOpt] Wrong code at -O1 since clang-18

Reproducer: https://godbolt.org/z/vGEz9nxMM
Reduced case:
```c
#include
#include
typedef uint32_t v4u32 __attribute__((vector_size(16)));
_Bool g5;
static void *g10;
v4u32 g26;
int main()
{
uint32_t bb3;
g10 = aligned_alloc(_Alignof(v4u32), sizeof(v4u32));
*(v4u32 *)g10 = g26;
__builtin_memmove(&bb3, g10, 4);
g5 = bb3;
}
```

This case triggers SIGSEGV at -O1, and terminates normally at -O0.

A quick look into the final IR at -O1:
```llvm
@g26 = dso_local local_unnamed_addr global <4 x i32> zeroinitializer, align 16
@g5 = dso_local local_unnamed_addr global i8 0, align 1
@g10.body = internal unnamed_addr global [16 x i8] undef, align 4

define dso_local noundef i32 @main() local_unnamed_addr {
entry:
%0 = load <4 x i32>, ptr @g26, align 16
store <4 x i32> %0, ptr @g10.body, align 16
%bb3.0.copyload = load i32, ptr @g10.body, align 16
%tobool = icmp ne i32 %bb3.0.copyload, 0
%storedv = zext i1 %tobool to i8
store i8 %storedv, ptr @g5, align 1
ret i32 0
}
```
It seems that the object returned by `aligned_alloc` is transformed into a global aligned with 4. But the accesses to it are still aligned with 16.