[Bug target/60410] [4.7/4.8/4.9 Regression] -fshort-double ICEs x86_64

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60410

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |rth at gcc dot gnu.org,
                   |                            |uros at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r189974, but that is just because it added an assertion.
The question is, if -fshort-double really has to be supported on x86_64/i686
(as ABI changing option), what to do with the intrinsics, grep double
config/i386/*intrin*.h shows quite a lot of intrinsics that use double, and
various builtins really assume DFmode arguments/return values/V*DF vectors etc.
So, perhaps we could just override
ix86_builtin_type_tab[(int)IX86_BT_DOUBLE] if double_type_node doesn't have
DFmode TYPE_MODE to some DFmode REAL_TYPE (that would likely fix this ICE), and
perhaps tweak the *intrin*.h headers to use typedef double __real_double
__attribute__((__mode__(__DFmode__))); instead of double (perhaps not this
part, it will very unlikely do the right thing, perhaps just tell people not to
use the intrinsics).