not have native support for this operation (such as X86). The legalized code uses two vector INT_TO_FP operations and is faster than scalarizing. llvm-svn: 127951