Hi!

On Tue, Feb 16, 2021 at 08:36:02PM +1100, Michael Ellerman wrote:
> Feng Tang <feng.t...@intel.com> writes:
> >   {standard input}:577: Error: unsupported relocation against base
> >   {standard input}:580: Error: unsupported relocation against base
> >   {standard input}:583: Error: unsupported relocation against base

> > The reason is macro 'mfdcr' requirs an instant number as parameter,
> > which is not met by show_plbopb_regs().
> 
> It doesn't require a constant, it checks if the argument is constant:
> 
> #define mfdcr(rn)                                             \
>       ({unsigned int rval;                                    \
>       if (__builtin_constant_p(rn) && rn < 1024)              \
>               asm volatile("mfdcr %0," __stringify(rn)        \
>                             : "=r" (rval));                   \
>       else if (likely(cpu_has_feature(CPU_FTR_INDEXED_DCR)))  \
>               rval = mfdcrx(rn);                              \
>       else                                                    \
>               rval = __mfdcr(rn);                             \
>       rval;})

It requires a constant number with known (at compile time) value, while
__builtin_constant_p checks for any constant.  The address of some
defined symbol is a constant as well normally, for example.

It's better to write that asm as
        asm volatile("mfdcr %0,%1" : "=r" (rval) : "n"(rn));
btw (the "n" constraint means "constant integer with known value" (it
stands for "numeric"), while the "i" constraint means just "constant
integer").

> But the error you're seeing implies the compiler is choosing the first
> leg of the if, even when rn == "base + x", which is surprising.
> 
> We've had cases in the past of __builtin_constant_p() returning false
> for things that a human can see are constant at build time, but I've
> never seen the reverse.

And it doesn't here :-)

But, you need some way to figure out an arg is a constant known number
here.  We don't have a builtin for that I think.  Maybe some trick can
be done?  Maybe simply test "rn >= 0" as well, does that work?


Segher

Reply via email to