On Mon, Feb 04, 2002 at 05:28:57PM -0800, H . J . Lu wrote: > On Mon, Feb 04, 2002 at 04:46:07PM +0000, Dominic Sweetman wrote: > > > > H . J . Lu (hjl@lucon.org) writes: > > > > > I can change glibc not to use branch-likely without using nop. But it > > > may require one or two instructions outside of the loop. Should I do > > > it given what we know now? > > > > I would not recommend using "branch likely" in assembler coding, if > > that's what you're asking. > > > > Here is a patch to remove branch likely. But I couldn't find a way > not to fill the delay slot with nop. BTW, is that safe to remove > ".set noreorder"? You mean, if there is nothing which can be put there? Yes, it's safe. > --- libc/linuxthreads/sysdeps/mips/pspinlock.c.llsc Mon Feb 4 13:45:01 2002 > +++ libc/linuxthreads/sysdeps/mips/pspinlock.c Mon Feb 4 17:09:02 2002 > @@ -40,7 +40,7 @@ __pthread_spin_lock (pthread_spinlock_t > "bnez %1,1b\n\t" > " li %2,1\n\t" > "sc %2,%0\n\t" > - "beqzl %2,1b\n\t" > + "beqz %2,1b\n\t" > " ll %1,%3\n\t" > ".set pop" > : "=m" (*lock), "=&r" (tmp1), "=&r" (tmp2) Is that really what you meant to do? The ll is now in the delay slot of the beqz. > --- libc/linuxthreads/sysdeps/mips/pt-machine.h.llsc Mon Feb 4 13:45:01 2002 > +++ libc/linuxthreads/sysdeps/mips/pt-machine.h Mon Feb 4 17:09:36 2002 > @@ -66,7 +66,7 @@ __compare_and_swap (long int *p, long in > " move %0,$0\n\t" > "move %0,%4\n\t" > "sc %0,%2\n\t" > - "beqzl %0,1b\n\t" > + "beqz %0,1b\n\t" > " ll %1,%5\n\t" > ".set pop\n" > "2:\n\t" Ditto. > @@ -89,7 +89,7 @@ compare_and_swap (volatile long int *p, > " move %0,$0\n\t" > "move %0,%4\n\t" > "sc %0,%2\n\t" > - "beqzl %0,1b\n\t" > + "beqz %0,1b\n\t" > " ll %1,%5\n\t" > ".set pop\n" > "2:\n\t" Ditto. -- Daniel Jacobowitz Carnegie Mellon University MontaVista Software Debian GNU/Linux Developer
