In both __giveup_fpu() and __giveup_altivec() we make two modifications
to tsk->thread.regs->msr. gcc decides to do a read/modify/write of
each change, so we end up with a load hit store:
        ld      r9,264(r10)
        rldicl  r9,r9,50,1
        rotldi  r9,r9,14
        std     r9,264(r10)
...
        ld      r9,264(r10)
        rldicl  r9,r9,40,1
        rotldi  r9,r9,24
        std     r9,264(r10)
Fix this by using a temporary.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
 #ifdef CONFIG_PPC_FPU
 void __giveup_fpu(struct task_struct *tsk)
 {
+       unsigned long msr;
+
        save_fpu(tsk);
-       tsk->thread.regs->msr &= ~MSR_FP;
+       msr = tsk->thread.regs->msr;
+       msr &= ~MSR_FP;
 #ifdef CONFIG_VSX
        if (cpu_has_feature(CPU_FTR_VSX))
-               tsk->thread.regs->msr &= ~MSR_VSX;
+               msr &= ~MSR_VSX;
 #endif
+       tsk->thread.regs->msr = msr;
 }
 
 void giveup_fpu(struct task_struct *tsk)
 
 static void __giveup_altivec(struct task_struct *tsk)
 {
+       unsigned long msr;
+
        save_altivec(tsk);
-       tsk->thread.regs->msr &= ~MSR_VEC;
+       msr = tsk->thread.regs->msr;
+       msr &= ~MSR_VEC;
 #ifdef CONFIG_VSX
        if (cpu_has_feature(CPU_FTR_VSX))
-               tsk->thread.regs->msr &= ~MSR_VSX;
+               msr &= ~MSR_VSX;
 #endif
+       tsk->thread.regs->msr = msr;
 }
 
 void giveup_altivec(struct task_struct *tsk)