On Jul 3, 2007, at 7:43 AM, Duncan Sands wrote: > Hi Dan, >> + // fold (add x, undef) -> undef >> + if (N1.getOpcode() == ISD::UNDEF) >> + return N1; > > what if N0 is undef and not N1?
Presumably, undef is being canonicalized to the RHS somewhere. I don't know if this is actually true in the dag combiner though, it should be checked. >> + // fold (mul x, undef) -> 0 >> + if (N1.getOpcode() == ISD::UNDEF) >> + return DAG.getConstant(0, VT); > > Yes, this seems better :) Again, is the undef guaranteed to be > in N1 and not N0? Does this work for vectors? Does DAG.getConstant(0, MVT::v4i32) work? This xform should probably be disabled after legalize for vectors. >> + // fold (or x, undef) -> -1 >> + if (N1.getOpcode() == ISD::UNDEF) >> + return DAG.getConstant(-1, VT); > > Is this the right way to get an all-bits-one value? Yes, but this has the same problem with vectors. :) -Chris _______________________________________________ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
