This is currently obscured by -Waddress warning disabled for Hotspot. If you enable this warning, then GCC would complain:

```
/home/shade/trunks/jdk/src/hotspot/share/opto/postaloc.cpp: In member function 'int PhaseChaitin::elide_copy(Node*, int, Block*, Node_List&, Node_List&, bool)':
/home/shade/trunks/jdk/src/hotspot/share/opto/postaloc.cpp:260:14: error: the compiler can assume that the address of 'value' will never be NULL [-Werror=address]
  260 | if( &value == NULL ) return blk_adjust;
      | ^
```

This thing is not theoretical, I played with the following snippet in Godbolt:

```
#include <cstddef>

class X {};

bool test(X& ref) {
    return (&ref == NULL);
}

int main() {
    return test(*((X*)NULL));
}
```

With -O2, GCC 5.5 and lower compiles `test` to actual check, like you would expect for pointers:

```
 test rdi,rdi
 sete al
 ret
```

With -O2, GCC 6.1 and higher compiles `test` to "return false", as warning told it might happen:

```
 xor eax,eax
 ret
```

Plus, adding these asserts to Hotspot:

```
diff --git a/src/hotspot/share/opto/postaloc.cpp b/src/hotspot/share/opto/postaloc.cpp
index 96c30a122bb..23b78305af2 100644
--- a/src/hotspot/share/opto/postaloc.cpp
+++ b/src/hotspot/share/opto/postaloc.cpp
@@ -536,4 +536,6 @@ void PhaseChaitin::post_allocate_copy_removal() {
       // Remove copies along phi edges
       for (uint k = 1; k < phi_dex; k++) {
+ assert(blk2value[pb->_pre_order] != nullptr, "UB");
+ assert(blk2regnd[pb->_pre_order] != nullptr, "UB");
         elide_copy(block->get_node(k), j, block, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false);
       }
```

Fails immediately during the build, so we do experience nulls here.