[PATCH] Resolve issues with fmuladd intrinsic handling across multiple backends

[Stephen Lin]

> Also, I took this opportunity to modify DAGCombiner to only check for
> type legality when forming ISD::FMA nodes (in unsafe math mode) after
> legalization, since this is allowed by the new function interface
> definition (as long as isFMAFasterThanFMulAndFAdd returns true for the
> type).  In practice with current target implementations this change
> seems to be a no-op, since the FMA will be formed after legalization
> anyway; however, there doesn't seem to be any harm in forming the FMA

In addition to the replies from Hal, Owen, and Tim on the list, Nadav
has said this is OK as well, off-list.

This does invoke an API break for out-of-tree targets, though, becomes
of the renaming above, but I think it's better this way because the
current function is broken and does not obey the interfaces stated.

Please let me know if there are any objections by tomorrow morning; I
plan on committing this then.

Stephen