x86-32 / x86-64 polyglot machine-code fragment that detects 64bit mode at run-time?

Question

Is it possible for the same bytes of machine code to figure out whether they\'re running in 32 or 64 bit mode, and then do different things?

i.e. write polyglot mach

Answer 1

The easiest way is by using the one-byte inc opcodes that are repurposed as REX prefixes in 64bit mode. A REX prefix has no effect on jcc, so you can do:

xor    eax,eax       ; clear ZF
db  0x40             ; 32bit: inc eax.   64bit: useless REX prefix
jz   .64bit_mode     ; REX jcc  works fine

See also a 3-way polyglot that returns 16, 32, or 64 according to the mode it executes in: Determine your language's version on codegolf.SE.

---

Reminder: normally you don't want this as part of a compiled binary. Detect mode at build time so any decision based on this can optimize away instead of being done at runtime. e.g. with #ifdef __x86_64__ and/or sizeof(void*) (but don't forget that the ILP32 x32 ABI has 32-bit pointers in long mode).

---

Here's a full Linux/NASM program that uses syscall to exit(1) if run as 64bit, or int 0x80 to exit(0) if run as 32bit.

The use of BITS 32 and BITS 64 ensure that it assembles to the same machine code either way. (And yes, I checked with objdump -d to show the raw machine-code bytes)

Even so, I used db 0x40 instead of inc eax, to make it clearer what's special.

BITS 32
global _start
_start:
        xor    eax,eax          ; clear ZF
        db 0x40                 ; 32bit: inc eax.  64bit: useless REX prefix
        jz      .64bit_mode     ; REX jcc  still works

        ;jmp .64bit_mode   ; uncomment to test that the 64bit code does fault in a 32bit binary

.32bit_mode:
        xor     ebx,ebx
        mov     eax, 1          ; exit(0)
        int     0x80


BITS 64
.64bit_mode:
        lea  rdx, [rel _start]      ; An instruction that won't assemble in 32-bit mode.
        ;; arbitrary 64bit code here

        mov  edi, 1
        mov  eax, 231    ;  exit_group(1).
        syscall          ; This does SIGILL if this is run in 32bit mode on Intel CPUs

;;;;; Or as a callable function:
BITS 32
am_i_32bit:  ;; returns false only in 64bit mode
        xor     eax,eax

        db 0x40                 ; 32bit: inc eax
                                ; 64bit: REX.W=0
        ;nop                     ; REX nop  is  REX xchg eax,eax
        ret                     ; REX ret works normally, too

---

Tested and working. I build it twice to get different ELF metadata around the same machine code.

$ yasm -felf64 -Worphan-labels -gdwarf2 x86-polyglot-32-64.asm && ld -o x86-polyglot.64bit x86-polyglot-32-64.o
$ yasm -felf32 -Worphan-labels -gdwarf2 x86-polyglot-32-64.asm && ld -melf_i386 -o x86-polyglot.32bit x86-polyglot-32-64.o
$ ./x86-polyglot.32bit && echo 32bit || echo 64bit
32bit
$ ./x86-polyglot.64bit && echo 32bit || echo 64bit
64bit

(build commands from Assembling 32-bit binaries on a 64-bit system (GNU toolchain), linked from the FAQ section in the x86 tag wiki).