Re: [PATCH] Add VEC_COND_EXPR node

[Roger Sayle <roger at eyesopen dot com>]

On Tue, 21 Sep 2004, Richard Henderson wrote:
> On Thu, Sep 16, 2004 at 09:47:49AM -0700, Devang Patel wrote:
> > + DEFTREECODE (VEC_COND_EXPR, "cond_expr", 'e', 3)
>                                  ^^^
>                                  typo.
>
> Otherwise ok.

Hi Devang,

Would it possible to add some documentation for vec_cond_expr to
c-tree.texi?  I'm a bit concerned about the change of sematics for
the Boolean expressions that serve as the initial operand of ?:.
For example, previously Boolean expression always returned a
scalar Boolean value but in the new scheme they return the equivalent
of a vector of Bool.

Hence if A = VEC_COND_EXPR ( X < Y, B, C )

means  for(i=0; i<N; i++) A[i] = X[i] < Y[i] ? B[i] : C[i];

then we'll run into problems with EQ_EXPR and NE_EXPR when X and Y
are vector constants.  Equality and inequality is already defined
for vectors, even though the other relational operators aren't.

  A = VEC_COND_EXPR (X == Y, B, C)

probably intends to mean:

  for(i=0; i<N; i++) A[i] = X[i] == Y[i] ? B[i] : C[i];

but if tree-ssa propagates VECTOR_CSTS into the equality expression,
we may constant fold the equality operator to false, based upon whether
any of the elements of VECTOR_CST X differs from VECTOR_CST Y.

  A = VEC_COND_EXPR (false, B, C)

which ultimately, one would hope would be folded to A = C.



My apologies if these issues were considered and resolved when this
topic was discussed on the gcc list, but unfortunately keeping up
with gcc-patches is hard enough.  Has anyone considered VEC_LT_EXPR,
VEC_GE_EXPR, VEC_EQ_EXPR, etc...?

Roger
--