Re: [patch] sbitmap.h: Speed up EXECUTE_IF_SET_IN_SBITMAP.

[Jeffrey A Law <law@redhat.com>]

On Tue, 2004-10-19 at 20:10, Kazu Hirata wrote:
> Hi Richard,
> 
> > >     SBITMAP_ELT_TYPE *ptr_ = (SBITMAP)->elms;				\
> > >   									\
> > > !   for (word_num_ = (MIN) / (unsigned int) SBITMAP_ELT_BITS;		\
> > > !        word_num_ < size_; word_num_++, bit_num_ = 0)		\
> > >       {								\
> > > !       SBITMAP_ELT_TYPE word_ = ptr_[word_num_];			\
> > [...]
> > > +   SBITMAP_ELT_TYPE word_ = ptr_[word_num_] >> bit_num_;		\
> > 
> > Previously, setting MIN arbitrarily large would not read from PTR;
> > now it does.  I think this is a faulty translation of this function.
> 
> oops.  I found another problem of accessing PTR beyond its end.  Note
> that I check for the end condition *after* I increment word_num and
> get a new word_ like so.
> 
>   for (;								\
>        word_num_ < size_;						\
>        word_num_++, bit_num_ = 0, word_ = ptr_[word_num_])		\
> 
> I'll be testing the patch below shortly.  The end result is a bit
> ugly, but there shouldn't be any performance loss.  OK to apply if
> testing passes?
> 
> Kazu Hirata
> 
> 2004-10-19  Kazu Hirata  <kazu@cs.umass.edu>
> 
> 	* sbitmap.h (EXECUTE_IF_SET_IN_SBITMAP): Don't access PTR
> 	beyond its end.
This is OK assuming it passed regression testing.
jeff