Re: Problems with a patch to improve dependence analysis

[Zdenek Dvorak <rakdver at kam dot mff dot cuni dot cz>]

Hello,

> > this condition is not restrictive enough; try using
> > 
> > (EXPR_P (def_stmt_rhs) && !REFERENCE_CLASS_P (def_stmt_rhs))
> > 
> 
> Yes, this seems to have fixed my problem. :) Why is this restriction
> nessecary?

references also match EXPR_P.  If you have expression like

p_1 = a.x;

loop
  {
    *p1 = ...
  }

you would propagate a.x to the base of the memory reference; this is
incorrect, as we do not know whether a.x is invariant in the loop.
Also, if you emit this expression containing memory references, you need
to update ssa form for virtual operands, which is why you have seen the
ICE.

Zdenek