This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] optabs and tree-codes for vector operations


On Aug 14, 2004, at 11:57 PM, Richard Henderson wrote:

So we'd have

a = VEC_COND_EXPR <(y < z), b, c>

Assume a[i] accesses vector element i. Then this would mean

forall i: a[i] = y[i] < z[i] ? b[i] : c[i]

The important consideration here is that we don't care how the target
implments the comparison, because we never expose it.

OK.

As for the rtl-level representation, see how mov<mode>cc is
implemented.  I think that's a better model (as it uses only
one named pattern) than conditional branches (which use hordes
of named patterns).

So, we'd have one new optab to for VEC_COND_EXPR. And we use target hooks to expand appropriately based on target requirements?

-
Devang

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]