Re: [PATCH 2/5][P1][tree-optimization/71437] Record more equivalences in some cases

[Jeff Law <law at redhat dot com>]

On 03/16/2017 06:54 AM, Trevor Saunders wrote:
> > +	  if (TREE_CODE (op0) == SSA_NAME && integer_zerop (op1))
> > +	    {
> > +	      enum tree_code code = eq->cond.ops.binary.op;
> > +	      if ((code == EQ_EXPR && eq->value == boolean_true_node)
> > +		  || (code == NE_EXPR && eq->value == boolean_false_node))
> > +		derive_equivalencs_from_bit_ior (op0, const_and_copies);
> > +
> > +	      /* TODO: We could handle BIT_AND_EXPR in a similar fashion
> > +		 recording that the operands have a nonzero value.  */
> > +
> > +	      /* TODO: We can handle more cases here, particularly when OP0 is
> > +		 known to have a boolean range.  */
>
> I don't think its necessarily useful to put a list here of all possible
> improvements, but we could also handle things like if ((a | b) </> 0)
> since those imply !=.
The equivalences you derive here must hold for both objects.  So without 
additional range information you can't really determine anything about 
(a | b) > 0.

If a and b are booleans or unsigned, then ideally we'd canonicalize the 
test into an [in]equality test.  At which point the existing code would 
fire recording  a = 0 and b = 0 on the false edge.  You still can't 
record anything on the true edge.  I don't think we're good at the 
canonicalization these days.

For (a & b) != 0, unconstrained we could record a != 0 and b != 0 on the 
true edge, but nothing can be recorded on the false edge.

If we know a and b have boolean ranges we can refine that to a = 1 and b 
= 1 on the true edge, but again, nothing can be recorded on the false edge.



Jeff