Problem with inline assembly and -O2 or -O3

[Jeremy Friesner]

Hi all,

I have a problem with some simple inline assembly -- I wrote a function to byte 
swap a 64-bit integer under x86, but it only works when compiled with -O0 or 
-O1.  When compiled with -O2 or -O3, it gives the wrong result.  Is this a 
compiler bug or have I done something wrong?

I'm using gcc/g++ 3.3.3 on a dual P3 under SUSE 9.1, FWIW.

Here is the correct output I get when compiling with -O0 or -O1:

$ ./swap
Test completed successfully.

and here is the (unexpected) output I get when compiling with -O2 or -O3:

$ ./swap
Error!  backagain=0 orig=1

And finally, here is the test program that demonstrates the problem... can 
anyone give me a suggestion?

--------------------------------------------------------------

#include <stdio.h>

typedef unsigned long long uint64;
typedef unsigned long uint32;

inline uint64 Swap64(uint64 val)
{
   uint64 a;
   volatile uint32 * inLo  = (uint32 *) &val;
   volatile uint32 * outLo = (uint32 *) &a;
   volatile uint32 * inHi  = (inLo+1);
   volatile uint32 * outHi = (outLo+1);
   __asm__ __volatile__ ("bswap %0" : "=r" (*inLo) : "0" (*inLo));
   __asm__ __volatile__ ("bswap %0" : "=r" (*inHi) : "0" (*inHi));
   *outHi = *inLo;
   *outLo = *inHi;
   return a;
}

int main(int, char **)
{
   for (uint64 x=0; x<1000; x++)
   {
      uint64 swapped = Swap64(x);
      uint64 backagain = Swap64(swapped);
      if (backagain != x)
      {
         printf("Error!  backagain=%llu orig=%llu\n", backagain, x);
         return 10;
      }
   }
   printf("Test completed successfully.\n");
   return 0;
}