Re: [tree-ssa] copy propagation and the abstraction penalty

[Joe Buck <jbuck@synopsys.com>]

On Thu, May 15, 2003 at 03:17:48PM -0700, Richard Henderson wrote:
> On Thu, May 15, 2003 at 10:10:22AM -0400, Diego Novillo wrote:
> >         I'm also wondering if we could change the
> >         may-alias between this and UV2150 to a must-alias, which would
> >         completely free this program from aliasing problems:
> 
> Indeed.  And for this case I definitely think it's the right thing to do.
> 
> IMO constant propagation should be able to take
> 
>     T.8_2 = &<UVa150>;
>     {
>       struct complex * const this;
> 
>       this_3 = (struct complex * const)T.8_2;
>       {
>         this->re = 1.0e+0;
> 
> and turn it into
> 
> 	(&<UVa150>)->re = 1.0e+0
> 
> which folds to
> 
> 	<UVa150>.re = 1.0e+0
> 
> At which point we have no aliasing problem, and a subsequent round
> of constant propagation ought to be able to send 1.0e+0 to its 
> destination.

That helps this case, but in many other cases the content of the temporary
struct's field will be a variable, and we would still want to copy-propagate.
For example, consider bit vector classes.  Typically the [] operator will
be overloaded to return a "bitref" object, which is a struct that has two
fields: a reference to the bit vector, and a bit offset.  The compiler
should be able to eliminate the temporary object.  We would have

struct bitref;

class bitvec {
public:
    void set_bit(unsigned pos, bool value);
    bool get_bit(unsigned pos) const;
    inline bitref operator[](unsigned pos);
};

struct bitref {
    bitref(bitvec& o, unsigned p) : obj(o), pos(p) {}
    bitvec& obj;
    unsigned pos;
    operator bool() const { return obj.get_bit(pos);}
    void operator=(bool value) { obj.set_bit(pos, value);}
    void operator=(const bitref& src) { obj.set_bit(pos, src);}
};

inline bitref bitvec::operator[](unsigned pos) { return bitref(*this, pos);}

and we want to compile

void assign(bitvec& dest, bitvec& src, unsigned i, unsigned j) {
    dest[i] = src[j];
}

Ideally, we should be able to do copy propagation good enough to turn this
into

    dest.set_bit(i, src.get_bit(j));

which means that we can kill the two generated bitref objects.  Note,
though, that there are no constants.  We have {&dest,i} and {&src,j}.

This kind of thing occurs throughout common C++ codes, and really kills us
on the Boost graph library.