Re: small improvement for fill and fill_n

[Dan Nicolaescu <dann at godzilla dot ics dot uci dot edu>]

Paolo Carlini <pcarlini@suse.de> writes:

  > Hi Dan,

Hi!

  > > Note that the memory location pointed to by __value is accessed in
  > > each loop iteration, the compiler cannot prove that __first and
  > > __value do not alias. One solution would be to use a temporary like
  > > so:
  > 
  > First, thanks for the analysis: in order to seriously pursue performance
  > improvements we really need this kind of work.

Well, this one was just a lucky observation while I was looking for
something unrelated. 

  > Coming to the point, however, I wonder whether there are hopes that the
  > alias analysis will be improved in a not-to-far future... It would be
  > so sad to add those temporaries: I suspect *many* would be needed in
  > the whole libraries... .-(

Diego was saying before the tree-ssa merge that he wants to work on
aliasing, so you might want to ask him. 

Another issue for C++ is that SRA does not work when the objects
involved use inheritance. 

Now for fill and fill_n, I don't think that any improvements to
aliasing can help given the way these functions are currently
written. There is not guarantee that __value is not a reference to one
of the objects that fill[n] iterate over. So IMHO using a temporary
is unavoidable for these 2 functions if the above assumption holds.

        --dan