Help! asm problems with 2.95 and 2.7.2

[Andrew Lees <andrewl at acay dot com dot au>]

Hello,

I am trying to construct a cross compiler from intel-glibc2.1 to intel-libc5 
for embedded system construction, with both host and target using 2.2.14 
linux kernels.  I have tried both gcc-2.95 and gcc-2.7.2.  In both cases I 
get errors when compiling the libc5 library, specifically compiling some 
cases of syscall* for the *shared* library build only.  The static library 
compiles fine.  I have confirmed that the compiler option causing the problem 
is -fPIC.  Optimisation options have no effect.  The error messages with both 
compiler versions is a variant of:

__adjtime.c: In function `adjtimex':
__adjtime.c:67: fixed or forbidden register was spilled.
This may be due to a compiler bug or to impossible asm
statements or clauses.

The code actually being compiled is:

inline static
 int    adjtimex (  struct timex *    ntx ) { long __res; __asm__ __volatile  
("int $0x80" : "=a" (__res) : "0" (124  ),"b" ((long)(  ntx ))); do { if 
((unsigned long)( __res ) >= (unsigned long)(-125)) { (*__errno_location ())  
= -( __res );  __res  = -1; } return (  int  ) ( __res ); } while (0) ; } ;

The error occurs when the inline is used in the calling function.

The error message given above was from 2.7.2.  2.95.2 is specific that 
register bx is the problem, yet it is not being used in an output context, 
nor is it marked as clobbered in the asm macro.  The asm macro used is that 
from the 2.2.14 kernel rather than that originally provided with the libc5 
tree - it seems safer to use the up-to-date syscall macros.

This occurs when using the first version of the cross (compiled with 
glibc2.1).

I have read the FAQ re asm register problems, (hence the attempt with 2.7.2) 
but that doesn't seem to apply directly to this problem.

The problem occurs even if the syscall1 macro is extracted from the code 
where it is currently used and placed inside a trivial function with no other 
register variables.

Any suggestions for a fix?  It seems strange that gcc won't compile syscall 
code.

Thanks for any help,

Andrew Lees.