We emit an error when -msoft-float and -maltivec/-mvsx is used together,
but when -msoft-float is used on its own, there is still +altivec and
+vsx in the IR attributes. This patch disables altivec and vsx and all
related sub features when -msoft-float is used.
