On Wed, Dec 14, 2022 at 06:11:26PM +0800, Kewen.Lin wrote:
> > The hacks with different precisions of powerpc 128-bit floating types are
> > very unfortunate, it is I assume because the middle-end asserted that scalar
> > floating point types with different modes have different precision.
> > We no longer assert that, as BFmode and HFmode (__bf16 and _Float16) have
> > the same 16-bit precision as well and e.g. C++ FE knows to treat standard
> > vs. extended floating point types vs. other unknown floating point types
> > differently in finding result type of binary operations or in which type
> > comparisons will be done.
>
> It's good news, for now those three long double modes on Power have different
> precisions, if they can have the same precision, I'd expect the ICE should be
> gone.
I'm talking mainly about r13-3292, the asserts now check different modes
have different precision unless it is half vs. brain or vice versa, but
could be changed further, but if the precision is the same, the FEs
and the middle-end needs to know how to deal with those.
For C++23, say when __ibm128 is the same as long double and _Float128 is
supported, the 2 types are unordered (neither is a subset or superset of
the other because there are many _Float128 values one can't represent
in double double (whether anything with exponent larger than what double
can represent or most of the more precise values), but because of the
variable precision there are double double values that can't be represented
in _Float128 either) and so we can error on comparisons of those or on
arithmetics with such arguments (unless explicitly converted first).
But for backwards compatibility we can't do that for __float128 vs. __ibm128
and so need to backwards compatibly decide what wins. And for the
middle-end say even for mode conversions decide what is widening and what is
narrowing even when they are unordered.
Jakub
