Editions Ideas

[Timon Gehr]

On 12/18/25 06:04, Walter Bright wrote:
> On 12/14/2025 8:02 PM, Timon Gehr wrote:
>> To be fair, D does have a broken floating point implementation.
> 
> Only insofar as the x87's implementation.
> ...

https://dlang.org/spec/float.html

> 
>> Which then by the D spec does not matter because D is actually allowed 
>> to randomly use the wrong floating point types and introduce 
>> additional conversions and rounding imprecision.
> 
> This was fixed a few years ago,

https://dlang.org/spec/float.html

> excepting that we do what everyone else does with the x87.
> ...

This can't really be true, because there is not even a single x87 
behavior that is used.
https://github.com/dlang/dmd/issues/20339

We might or might not use the x87 even when there is no `real` involved 
anywhere and we are working strictly with `float` or `double` data. 
Neither `float` nor `double` arithmetic is even supported by the x87 
unless you configure it to not do double rounding, and you have to write 
back to memory after every operation to catch out-of-range exponents. 
(Afaiu none of the D implementations do any of this.)
https://dlang.org/spec/float.html

> 
>> Here's another recent complaint:
>> https://forum.dlang.org/thread/hvftxmrqvrmficcvlkqi@forum.dlang.org
> 
> Different compilers have different long double sizes.
> ...

Not the problem, D does not have "long double", and we do not even have 
to use `real` to get the broken behavior. If you read that thread, 
obviously equivalent D code that executes the same float operations on 
the same declared types (`double`!) in the same order works correctly 
only when the optimizer happens to feel like it. (This is comparing D 
code to D code, no "long double" or `real` or any other 
implementation-defined floating-point storage type involved.)

> None of that has anything to do with NaN behavior.

The justification of not wanting to have a broken floating point 
implementation applies in all of these cases.

Anyway, this is on topic, because changing e.g., constant folding to 
work correctly can be a breaking change.