Re: [GIT PULL] optimize 64-by-32 ddivision for constant divisors on 32-bit machines

[Nicolas Pitre]

On Mon, 23 Nov 2015, Arnd Bergmann wrote:

> On Monday 23 November 2015 11:04:33 Nicolas Pitre wrote:
> > 
> > OK... I'm able to "fix" the build with:
> > 
> > diff --git a/include/asm-generic/div64.h b/include/asm-generic/div64.h
> > index 163f77999e..d246c4c801 100644
> > --- a/include/asm-generic/div64.h
> > +++ b/include/asm-generic/div64.h
> > @@ -206,7 +206,7 @@ extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
> >         uint32_t __rem;                                 \
> >         (void)(((typeof((n)) *)0) == ((uint64_t *)0));  \
> >         if (__builtin_constant_p(__base) &&             \
> > -           is_power_of_2(__base)) {                    \
> > +           is_power_of_2(__base) && __base != 0) {     \
> >                 __rem = (n) & (__base - 1);             \
> >                 (n) >>= ilog2(__base);                  \
> >         } else if (__div64_const32_is_OK &&             \
> > 
> > What doesn't make sense to me is the fact that is_power_of_2() is 
> > defined as:
> > 
> > static inline __attribute__((const))
> > bool is_power_of_2(unsigned long n)
> > {
> >         return (n != 0 && ((n & (n - 1)) == 0));
> > }
> > 
> > So the test for zero is already in there.
> > 
> > And adding BUILD_BUG_ON(__builtin_constant_p(__base) && __base == 0) 
> > before the if doesn't trig either.
> 
> I've seen similarly messed up situations with PROFILE_ALL_BRANCHES
> before, I think it's got something to do with how __builtin_constant_p()
> is used inside of the __trace_if() macro, and how gcc sometimes falls
> back to treating variables as not-really-constant based on context.
> 
> To gcc, __builtin_constant_p is just best-effort, and they don't care
> about returning false sometimes if they catch most cases in practice.

But here it must have returned true, and is_power_of_2() returned true 
as well (which implies that __base is not zero), ans somehow aving an 
additional __base != 0 test changes the outcome.  There is a correctness 
issue beyond __builtin_constant_p it seems.

> Note that llvm will always return false for __builtin_constant_p on
> non-pointer arguments, which breaks a lot of optimizations.

If llvm is able to optimize this case on its own then we won't need all 
this contraption.


Nicolas
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/