yann@1: See http://gcc.gnu.org/PR15089
yann@1: "In some cases the specified register for a local variable meant to be
yann@1: used with inline assembly code is not respected.  This breaks code 
yann@1: relying on this feature to implement non-standard calling convension."
yann@1: 
yann@1: This fix is probably needed to properly implement linux syscalls in some architectures.
yann@1: 
yann@1: First added to crosstool by Jamie Hicks, see
yann@1: http://www.handhelds.org/hypermail/kernel-discuss/current/0066.html
yann@1: 
yann@1: Index: gcc/loop.c
yann@1: ===================================================================
yann@1: RCS file: /cvs/gcc/gcc/gcc/loop.c,v
yann@1: retrieving revision 1.488.2.3
yann@1: diff -u -r1.488.2.3 loop.c
yann@1: --- gcc-old/gcc/loop.c	14 Feb 2004 14:46:03 -0000	1.488.2.3
yann@1: +++ gcc-new/gcc/loop.c	28 Apr 2004 22:02:53 -0000
yann@1: @@ -929,6 +929,7 @@
yann@1:  			  || (! (GET_CODE (SET_SRC (set)) == REG
yann@1:  				 && (REGNO (SET_SRC (set))
yann@1:  				     < FIRST_PSEUDO_REGISTER))))
yann@1: +		      && regno >= FIRST_PSEUDO_REGISTER 
yann@1:  		      /* This test is not redundant; SET_SRC (set) might be
yann@1:  			 a call-clobbered register and the life of REGNO
yann@1:  			 might span a call.  */