ELF Shared Object in x86-64 Assembly language

I've already compiled with -fPIC, so what's wrong?

That part of the error message is for people who are linking compiler-generated code.

You're writing asm by hand, so as datenwolf correctly wrote, when writing a shared library in assembly, you have to take care for yourself that the code is position independent.

This means file must not contain any 32-bit absolute addresses (because relocation to an arbitrary 64-bit base is impossible). 64-bit absolute relocations are supported, but normally you should only use that for jump tables.

---

mov var1, %rcx uses a 32-bit absolute addressing mode. You should normally never do this, even in position-dependent x86-64 code. The normal use-cases for 32-bit absolute addresses are: putting an address into a 64-bit register withmov $var1, %edi (zero-extends into RDI)
and indexing static arrays: mov arr(,%rdx,4), %edx

mov var1(%rip), %rcx uses a RIP-relative 32-bit offset. It's the efficient way to address static data, and compilers always use this even without -fPIE or -fPIC for static/global variables.

You have basically two possibilities:

• Normal library-private static data, like C compilers will make for __attribute__((visibility("hidden"))) long var1;, same as for -fno-PIC.

  .data
      .globl var1       # linkable from other .o files in the same shared object / library
      .hidden var1      # not visible for *dynamic* linking outside the library
  var1:
      .quad     0x012345
  
  .text
      .globl func1
  func1:
      xor  %eax, %eax             # return 0
      mov  var1(%rip), %rcx   
      ret
  

• full symbol-interposition-aware code like compilers generate for -fPIC.

  You have to use the Global Offset Table. This is how a compiler does it, if you tell him to produce code for a shared library. Note that this comes with a performance hit because of the additional indirection.

  See Sorry state of dynamic libraries on Linux for more about symbol-interposition and the overheads it imposes on code-gen for shared libraries if you're not careful about restricting symbol visibility to allow inlining.

  var1@GOTPCREL is the address of a pointer to your var1, the pointer itself is reachable with rip-relative addressing, while the content (the address of var1) is filled by the linker during loading of the library. This supports the case where the program using your library defined var1, so var1 in your library should resolve to that memory location instead of the one in the .data or .bss (or .text) of your .so.

      .section .data
      .globl var1
      # without .hidden
  var1:
      .quad     0x012345
  
      .section .text
      .globl func1
  func1:
      xor %eax, %eax
      mov var1@GOTPCREL(%rip), %rcx
      mov (%rcx), %rcx
      ret
  

See some additional information at http://www.bottomupcs.com/global_offset_tables.html

An example on the Godbolt compiler explorer of -fPIC vs. -fPIE shows the difference that symbol-interposition makes for getting the address of non-hidden global variables:

• movl $x, %eax 5 bytes, -fno-pie
• leaq x(%rip), %rax 7 bytes, -fPIE and hidden globals or static with -fPIC
• y@GOTPCREL(%rip), %rax 7 bytes and a load instead of just ALU, -fPIC with non-hidden globals.

Actually loading always uses x(%rip), except for non-hidden / non-static vars with -fPIC where it has to get the runtime address from the GOT first, because it's not a link-time constant offset relative to the code.

Related: 32-bit absolute addresses no longer allowed in x86-64 Linux? (PIE executables).

---

A previous version of this answer stated that the DATA and BSS segments could move relative to TEXT when loading a dynamic library. This is incorrect, only the library base address is relocatable. RIP-relative access to other segments within the same library is guaranteed to be ok, and compilers emit code that does this. The ELF headers specify how the segments (which contain the sections) need to be loaded/mapped into memory.

I don't understand. I've already compiled with -fPIC, so what's wrong?

-fPIC is a flag concerning the creation of machine code from non-machine code, i.e. which operations to use. In the compilation stage. Assembly is not compiled, though! Each assembly mnemonic maps directly to a machine instruction, your code is not compiled. It's just transcribed into a slightly different format.

Since you're writing it in assembly, your assembly code must be position independent to be linkable into a shared library. -fPIC has not effect in your case, because it only affects code generation.