It's not a new behaviour. First it was introduced by
2d5372cf
("lower -O2 optimization down to -O1 on UNREG") to fix build
failure on unregisterised powerpc64.

This time I've noticed build failures on unregisterised ia64.

The change was accidentally reverted by commit
14d0f7f1


("Build system: Add stage specific SRC_HC_(WARNING_)OPTS)

The revert happened due to the following code rearrangement:

    ifeq "$(GhcUnregisterised)" "YES"
    GhcStage1HcOpts=
    GhcStage2HcOpts=
    GhcStage3HcOpts=
    endif

    GhcUnregisterised=@Unregisterised@

As a result 'ifeq' part has no effect.

The change moves 'ifeq' down to the very end of file
and adds a note it depends on the 'GhcUnregisterised' variable.

Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>