Missed optimization to or disjoint

[GCC bug 127013](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127013)
[Godbolt](https://godbolt.org/z/Gfq3PTPdq)

With -O2, llvm should optimize to or disjoint:

```c
#define TYPE_PREC 32
#define TYPE unsigned _BitInt(TYPE_PREC)

TYPE
src_add (TYPE x, TYPE y)
{
if ((x & y) != 0)
__builtin_unreachable ();

return (x + y);
}

TYPE
src_or (TYPE x, TYPE y)
{
if ((x & y) != 0)
__builtin_unreachable ();

return (x | y);
}

TYPE
src_xor (TYPE x, TYPE y)
{
if ((x & y) != 0)
__builtin_unreachable ();

return (x ^ y);
}
```

```llvm
define dso_local noundef i32 @src_add(i32 noundef %x, i32 noundef %y) local_unnamed_addr {
entry:
%and = and i32 %y, %x
%cmp.not = icmp eq i32 %and, 0
tail call void @llvm.assume(i1 %cmp.not)
%add = add i32 %y, %x
ret i32 %add
}

define dso_local noundef i32 @src_or(i32 noundef %x, i32 noundef %y) local_unnamed_addr {
entry:
%and = and i32 %y, %x
%cmp.not = icmp eq i32 %and, 0
tail call void @llvm.assume(i1 %cmp.not)
%or = or i32 %y, %x
ret i32 %or
}

define dso_local noundef i32 @src_xor(i32 noundef %x, i32 noundef %y) local_unnamed_addr {
entry:
%and = and i32 %y, %x
%cmp.not = icmp eq i32 %and, 0
tail call void @llvm.assume(i1 %cmp.not)
%xor = xor i32 %y, %x
ret i32 %xor
}

declare void @llvm.assume(i1 noundef) #1
```

Start with the C reproducer and the linked Godbolt example, comparing the generated LLVM IR for src_add, src_or, and src_xor at -O2. Trace the optimization handling llvm.assume and the disjoint bitwise condition; done means the three operations are optimized appropriately without changing the shown semantics.

Written by the indexing model from the issue text.