android ndk asm compile error: inconsistent operand constraints in an 'asm'

Issue

I am compiling a piece of asm code for android:

static void MyCPUID(UInt32 function, UInt32 *a, UInt32 *b, UInt32 *c, UInt32 *d)
{
      __asm__ __volatile__ (
      "cpuid"
      : "=a" (*a) ,
      "=b" (*b) ,
      "=c" (*c) ,
      "=d" (*d)
      : "0" (function)) ;
}

The APP_ABI is set to 'all': APP_ABI := all

The compiling failed when come to x86:

$ ndk-build
[armeabi-v7a] Gdbserver      : [arm-linux-androideabi-4.6] libs/armeabi-v7a/gdbserver
[armeabi-v7a] Gdbsetup       : libs/armeabi-v7a/gdb.setup
[armeabi] Gdbserver      : [arm-linux-androideabi-4.6] libs/armeabi/gdbserver
[armeabi] Gdbsetup       : libs/armeabi/gdb.setup
[x86] Gdbserver      : [x86-4.6] libs/x86/gdbserver
[x86] Gdbsetup       : libs/x86/gdb.setup
[mips] Gdbserver      : [mipsel-linux-android-4.6] libs/mips/gdbserver
[mips] Gdbsetup       : libs/mips/gdb.setup
[armeabi-v7a] Compile thumb  : hello-jni <= CpuArch.c
[armeabi-v7a] SharedLibrary  : libhello-jni.so
[armeabi-v7a] Install        : libhello-jni.so => libs/armeabi-v7a/libhello-jni.so
[armeabi] Compile thumb  : hello-jni <= CpuArch.c
[armeabi] SharedLibrary  : libhello-jni.so
[armeabi] Install        : libhello-jni.so => libs/armeabi/libhello-jni.so
[x86] Compile        : hello-jni <= CpuArch.c
D:/adt/ndk/samples/hello-jni/jni/CpuArch.c: In function 'MyCPUID':
D:/adt/ndk/samples/hello-jni/jni/CpuArch.c:75:3: error: inconsistent operand constraints in an 'asm'
/cygdrive/d/adt/ndk/build/core/build-binary.mk:391: recipe for target '/cygdrive/d/adt/ndk/samples/hello-jni/obj/local/x86/objs-debug/hello-jni/CpuArch.o' failed
make: *** [/cygdrive/d/adt/ndk/samples/hello-jni/obj/local/x86/objs-debug/hello-jni/CpuArch.o] Error 1

I don't have much experience in asm. And the error msg seems not enough to find a solution. :(

BTW,the compiling is made in win7 using cygwin.

Full version:

static void MyCPUID(UInt32 function, UInt32 *a, UInt32 *b, UInt32 *c, UInt32 *d)
{
  #ifdef USE_ASM

  #ifdef _MSC_VER

  UInt32 a2, b2, c2, d2;
  __asm xor EBX, EBX;
  __asm xor ECX, ECX;
  __asm xor EDX, EDX;
  __asm mov EAX, function;
  __asm cpuid;
  __asm mov a2, EAX;
  __asm mov b2, EBX;
  __asm mov c2, ECX;
  __asm mov d2, EDX;

  *a = a2;
  *b = b2;
  *c = c2;
  *d = d2;

  #else

  __asm__ __volatile__ (
    "cpuid"
    : "=a" (*a) ,
      "=b" (*b) ,
      "=c" (*c) ,
      "=d" (*d)
    : "0" (function)) ;

  #endif

  #else

  int CPUInfo[4];
  __cpuid(CPUInfo, function);
  *a = CPUInfo[0];
  *b = CPUInfo[1];
  *c = CPUInfo[2];
  *d = CPUInfo[3];

  #endif
}

Solution

This code is based up something I wrote in this Stackoverflow answer. One has to be careful to preserve %ebx register on some x86 based architectures/ABI. %ebx is used to relocate code (shared object etc) when position independent code (-fPIC gcc option) is being generated. The code below avoids using =b in the extended assembler output and uses a register the compiler knows is free and usable. %ebx is preserved by swapping it to the free register before and after the call to cpuid. I've also fixed a small gotchya bug related to the %ecx register. I clear it to 0 ("c"(0)) since on some architectures failure to do so will result in stale values being returned by cpuid.

static void MyCPUID(UInt32 function, UInt32 *a, UInt32 *b, UInt32 *c, UInt32 *d)
{
#if defined(__i386__)
    __asm__ __volatile__ (
           "xchgl\t%%ebx, %k1\n\t"      \
           "cpuid\n\t"                  \
           "xchgl\t%%ebx, %k1\n\t"

      : "=a"(*a), "=&r"(*b), "=c"(*c), "=d"(*d)
      : "a"(function), "c"(0));

#elif defined(__x86_64__)
    __asm__ __volatile__ (
           "xchgq\t%%rbx, %q1\n\t"      \
           "cpuid\n\t"                  \
           "xchgq\t%%rbx, %q1\n\t"

      : "=a"(*a), "=&r"(*b), "=c"(*c), "=d"(*d)
      : "a"(function), "c"(0));
#else
#error "Unknown architecture."
#endif
}

Answered By - Michael Petch