On Wed, Jan 09, 2019 at 09:37:40AM +0000, Jonathan Wakely wrote:
> The IBM128 long double format isn't foldable in constant expressions, so
> conditionally skip the std::complex<long double> cases when they'll
> fail.
>
> PR libstdc++/88204
> * testsuite/26_numerics/complex/operators/more_constexpr.cc: Do not
> test std::complex<long double> if long double format is IBM128.
> * testsuite/26_numerics/complex/requirements/more_constexpr.cc:
> Likewise.
>
> Tested powerpc64le-linux, committed to trunk.
Thanks for doing that. Due to the variable number of bits in the mantissa
unless we represent the double-double constants in the compiler as a pair of
doubles there will be always many cases that can't be folded. That said,
I've been wondering if we couldn't fold at least the basic arithmetics on
some constants, essentially by emulating what will happen at runtime. I.e.
if we want to constant fold PLUS_EXPR (or MINUS_EXPR, MULT_EXPR, *DIV_EXPR,
and a few others) of two double-double REAL_CSTs, split them into two
double values by rounding to a double, subtracting that double and rounding
the remainder to double, perform whatever we emit inline or the runtime
library does for the basic operations and in the end check if the result of
that operation fits in the canonical format we pretend for the double-double
(i.e. the 106 bits mantissa etc.) and if it does, allow constant folding
that. E.g. in Red Hat bugzilla I think we have several spots where people
complain about this lack of folding, it breaks a couple of real-world apps
etc. Or is double-double considered dead format that will hopefully soon go
away, not worth spending more time on?
Jakub
