This is turned up in bug 114933 .  The problem is, that on i386 the syscallN  
macro is defined like so:  
  long __res; \  
  __asm__ volatile ("int $0x80" \  
        : "=a" (__res) \  
        : "0" (__NR_##name),"b" ((long)(arg1)),"c" ((long)(arg2)), \  
          "d" ((long)(arg3)),"S" ((long)(arg4)),"D" ((long)(arg5))); \  
  
If one of the arguments (in the _llseek syscall it's the arg4) is a pointer  
which the syscall is expected to write to (to the memory pointed to by this  
ptr), then this side-effect is not captured in the asm.  
  
If anyone uses this macro to define it's own version of the syscall  
(sometimes necessary when not using glibc) and it's inlined, then GCC  
doesn't know that this asm write to "*dest", when called like so for instance:  
  out = 1;  
  llseek (fd, bla, blubb, &out, trara)  
  use (out); 
Here nobody tells GCC that "out" actually is written to (just a pointer to it  
is passed to the asm).  Hence GCC might (and in the above bug did) 
copy-propagate "1" into the second use of "out". 
 
The easiest solution would be to add a "memory" clobber to the definition 
of this syscall macro.  As this is a syscall, it shouldn't inhibit too many 
optimizations.