unsigned long long notrace sched_clock(void)
 {
-       return clocksource_cyc2ns(read_c0_cvmcount(),
-                                 clocksource_mips.mult,
-                                 clocksource_mips.shift);
+       /* 64-bit arithmatic can overflow, so use 128-bit.  */
+#if (__GNUC__ < 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ <= 3))
+       u64 t1, t2, t3;
+       unsigned long long rv;
+       u64 mult = clocksource_mips.mult;
+       u64 shift = clocksource_mips.shift;
+       u64 cnt = read_c0_cvmcount();
+
+       asm (
+               "dmultu\t%[cnt],%[mult]\n\t"
+               "nor\t%[t1],$0,%[shift]\n\t"
+               "mfhi\t%[t2]\n\t"
+               "mflo\t%[t3]\n\t"
+               "dsll\t%[t2],%[t2],1\n\t"
+               "dsrlv\t%[rv],%[t3],%[shift]\n\t"
+               "dsllv\t%[t1],%[t2],%[t1]\n\t"
+               "or\t%[rv],%[t1],%[rv]\n\t"
+               : [rv] "=&r" (rv), [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3)
+               : [cnt] "r" (cnt), [mult] "r" (mult), [shift] "r" (shift)
+               : "hi", "lo");
+       return rv;
+#else
+       /* GCC > 4.3 do it the easy way.  */
+       unsigned int __attribute__((mode(TI))) t;
+       t = read_c0_cvmcount();
+       t = t * clocksource_mips.mult;
+       return (unsigned long long)(t >> clocksource_mips.shift);
+#endif
 }
 
 void __init plat_time_init(void)