On Saturday, 1 October 2016 at 19:10:47 UTC, Martin Nowak wrote:
Just tried to fix the float/double initialization w/ signaling
NaNs [¹], but it seems we can't reliably do that for all
backends/architectures. Any additional move of float might
convert SNaNs to QNaNs (quiet NaNs). This has also been the
finding of other people [²][³].
The biggest problem w/ the current situation is that float
fields of a struct sometimes are initialized to QNaNs and fail
`s.field is float.init`.
We thought about giving up on SNaNs as default float init
values. Is anyone relying on them?
I just had the same 'fun' with LDC. Both LDC 1.3 and DMD 2.074.0
produce special quiet NaNs for float.init and double.init on
Win64 (both most significant mantissa bits set). I also tried to
fix it, but it seems impossible when the x87 FPU (and not SSE) is
used. This leads to a Win64 LDC build using signalling inits when
cross-compiling via `-m32`, while the native Win32 LDC compiler
uses quiet ones etc.
So I'm all in for consistent special quiet NaNs as init values
for all 3 floating-point types (already implemented,
https://github.com/ldc-developers/ldc/pull/2207). If someone
relies on signalling NaNs and missed the original post, here's
your chance to speak up.
