We have a bad situation here as at the time of reload folding we can't
predict yet whether the spillslot coalescer must enlarge the spillslot
to 80bit precision.
For now I fix it by not folding reloads into x87 nodes. The better
alternative which I didn't do yet are:
- Move the spillslot coalescing to an earlier place before reload
  folding. This is somewhat tricky as the reload folding is implemented
  in generic code, while coalescing is backend specific.
- Revert folded AM when the spillslot is widened.