[tree-ssa] Out of SSA status and issues

[Andrew MacLeod]

On Tue, 2003-05-13 at 09:11, Diego Novillo wrote:
> On Tue, 2003-05-13 at 08:56, Michael Matz wrote:
> > Hi,
> > 
> > On 13 May 2003, Diego Novillo wrote:
> > 
> > > Why?
> > >
> > >      1. foo()
> > >      2. {
> > >      3.   int i, *p;
> > >      4.
> > >      5.   p = malloc();
> > >      6.   i = *p;
> > >      7.   return i + 9;
> > >      8. }
> > >
> > > I see nothing wrong in replacing 'i + 9' with '*p + 9'.
> > 
> > I do.  *p could have been changed in between 6 and 7.
> >
> The above is a complete program, *p has not changed between 6 and 7 nor
> it is volatile.
> 

but you are missing the point. There are lots of trivial examples which
dont show problems.

The problem occurs when yoiu add a new stmt, 6.5 *p = 10

copyprop still thinks it can propagate the stmt, but it can't.

*p = 10 is not a co[py. It is a store to memory, and you have to go
looking at aliasing and such in order to move around loads. copy prop
doesnt do that.

> > I.e. *p is not a register but an expression, and a copy instruction copies
> > registers to registers.
> > 
> Not in tree-ssa.  INDIRECT_REFs are first-class variables.  Since in
> GIMPLE pointers can only be one level deep, we can maintain SSA version
> numbers for the pointer and the pointed-to memory location.  Granted, it
> is somewhat painful (see my other message in this thread) and I'm
> starting to think that it may not be sufficiently worth the pain.
> 
> > > Not that tree-ssa will do anything with this code, the default
> > > type-based aliasing is too conservative, but PTA may disambiguate this.
> > 
> > Of course you _can_ optimize such cases, but it's not a subset of copy
> > propagation but, hmm, un-PRE or something like that.
> > 
> It is if INDIRECT_REFs are treated as regular variables, like we
> currently do.  I agree that otherwise you need to treat them as
> expressions.  If we stopped dealing with INDIRECT_REFs as first-class
> variables, we would probably move this in to the hands of SSAPRE.
> 

I think the only issue is that the copy prop pass is incorrectly
assuming that a load from memory through a pointer is the same thing as
a copy between variables, and it isn't.

Andrew