Hi, Maciej, On Wed, Apr 7, 2021 at 9:38 PM Maciej W. Rozycki <macro@xxxxxxxxxxx> wrote: > > On Wed, 7 Apr 2021, Huacai Chen wrote: > > > > This code is rather broken in an obvious way, starting from: > > > > > > unsigned long long __n; \ > > > \ > > > __high = *__n >> 32; \ > > > __low = __n; \ > > > > > > where `__n' is used uninitialised. Since this is my code originally I'll > > > look into it; we may want to reinstate `do_div' too, which didn't have to > > > be removed in the first place. > > I think we can reuse the generic do_div(). > > We can, but it's not clear to me if this is optimal. We have a DIVMOD > instruction which original code took advantage of (although I can see > potential in reusing bits from include/asm-generic/div64.h). The two > implementations would have to be benchmarked against each other across a > couple of different CPUs. The original MIPS do_div() has "h" constraint, and this is also the reason why Ralf rewrote this file. How can we reintroduce do_div() without "h" constraint? Huacai > > > > Huacai, thanks for your investigation! Please be more careful in > > > verifying your future submissions however. > > Sorry, I thought there is only one bug in div64.h, but in fact there > > are three... > > This just shows the verification you made was not good enough, hence my > observation. > > Maciej
