Bug 5974 – Incorrect nested function address in expression (&f).ptr

Created attachment 966 Problem curious test case Simple program: void main() { void f() { } void* t; t = (&f).ptr; //t = (&f).funcptr; //Uncommented: Error: &f is not an lvalue } First, look at lines 00402019 and 0040201E in disassembled main (dmd 1.067): void main() 00402010 enter 4,0 void* t; 00402014 xor eax,eax 00402016 mov dword ptr [t],eax t = (&f).ptr; 00402019 mov ecx,offset main@main@f (402028h) 0040201E mov dword ptr [t],eax 00402021 xor eax,eax //t = (&f).funcptr; //Uncommented: Error: &f is not an lvalue } 00402023 leave 00402024 ret 00402025 int 3 00402026 int 3 00402027 int 3 { void f() { } 00402028 enter 4,0 0040202C mov dword ptr [ebp-4],eax 0040202F leave 00402030 ret Second, why "&f is not an lvalue" error occurs for "(&f).funcptr"? Maybe, I don't understand something? Third, look at curiousTest.d attachment - it fails only if both asserts are uncommented.

More direct test case: void main() { void f() { } void* t1 = (&f).ptr; void* t2 = (&f).ptr; assert(t1 == t2); } The fac that (&f).funcptr fails to compile is a separate though closely related bug.

(In reply to comment #0) > Second, why "&f is not an lvalue" error occurs for "(&f).funcptr"? Maybe, I > don't understand something? iirc this fails because dmd rewrites dg.funcptr as *(cast(void**)(&dg)+1), which naturally doesn't work on an rvalue. The way to fix this is to rewrite it as cast(void*)(cast(uint<64/128>)dg >> <32/64>), the same way array length is accessed.

I believe this was fixed in https://github.com/dlang/dmd/pull/3181 - at least, the program in comment 2 stops failing after that change. Please reopen otherwise.