On Wed, Jun 15, 2016 at 03:12:55PM +0200, Richard Biener wrote: > On Wed, 15 Jun 2016, Michael Meissner wrote: > > Eventually, I decided to punt having to have explicit paths for widening. I > > used fractional modes for IFmode (ibm long double format) and KFmode (IEEE > > 128-bit format). IFmode widens to KFmode which widens to TFmode. A backend > > hook is used to not allow IBM long double to widen to IEEE long double and > > vice > > versa. At the moment, since there is no wider type than 128-bits, it isn't > > an > > issue. > > Ok, using fractional float modes is a lie though as both IFmode and KFmode > have no insignificant bits. I checked that if you have two same-size > FLOAT_MODEs they still get iterated over with GET_MODE_WIDER_MODE, > in order of appearance. So your reason to use fractional modes was > to make the order explicit and to avoid bogus handling in for example > the constant pool handling which would mistake say IFmode as being > smaller than KFmode just because one is GET_MODE_WIDER_MODE of the other > (and both have the same precision)? Interestingly for > GET_MODE_2XWIDER_MODE it "skips" the duplicate-sized and chooses > the "first" one. > > I guess having to have both float formats usable at the same time > is not really supported by our mode machinery. > > > Note, I do feel the front ends should be modified to allow __complex > > __float128 directly rather than having to use an attribute to force the > > complex type (and to use mode(TF) on x86 or mode(KF) on PowerPC). It > > would clean up both x86 and PowerPC. However, those patches aren't > > written yet. > > Sure, but that's independent of the modes issue. > > Anyway, I'm fine with the backport if you sort out the issue with > the assert in layout_type.
I did change the code as you suggested, and it did bootstrap and have no regressions. When I get back on Monday, I will look at implementing the fix in trunk, and then putting the revised patch into GCC 6.2. Thanks. -- Michael Meissner, IBM IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA email: meiss...@linux.vnet.ibm.com, phone: +1 (978) 899-4797
