used_math not cleared for new processes?

used_math not cleared for new processes?

[Zhang Fuxin]

hi,linux-mips,
   I find that current->used_math isn't cleared when we start a new process.Is it
intended? I mean 'start_thread' in do_exec but not 'copy_thread' in do_fork when
speaking 'start a new process'. We can/should? keep used_math for the latter,but for
the former?

   It leads to a failure in libm-test(the fpu control register doesn't equal to default)
I think the reason is that init_fpu fail to be called with used_math set.
    
  BTW: besides this failure,i see a lot of more related to extra "Invalid operation" exception.
e.g.:
   exp(NaN) == NaN: Exception "Invalid operation" set
   fmax (0, NaN) == 0: Exception "Invalid operation" set
..
 
  My cpu is IDT RC64474,which is basically a r4600.

  I think the cause probably is mips's special SNaN and QNaN handling.

  Could somebody explain?

Regards
            Zhang Fuxin
            fxzhang@ict.ac.cn

Re: used_math not cleared for new processes?

[Jun Sun]

Zhang Fuxin wrote:

> hi,linux-mips,
>    I find that current->used_math isn't cleared when we start a new process.Is it
> intended? I mean 'start_thread' in do_exec but not 'copy_thread' in do_fork when
> speaking 'start a new process'. We can/should? keep used_math for the latter,but for
> the former?
> 


I think it make sense to clear used_math in do_exec().  It also improvies the
 performance slightly by not loading the parent's FPU context when it uses the
FPU for the first time.

Do you have a patch for this?

Jun