[fpc-devel] Different computation precision in 32-bit and 64-bit app

[Sven Barth]

NetSpirit via fpc-devel <fpc-devel at lists.freepascal.org> schrieb am Fr.,
30. Juli 2021, 15:37:

> This code gives distinct results when compiled for Win32 and Win64
> - Win32: 49744125
> - Win64: 49744124.9999999
>
> procedure Test;
> var
>   D: Double;
> begin
>   D := 3 * (exp(3 * ln(255)));
>   WriteLn('Result: ' + FloatToStr(D));
> end;
>
>
> Somewhere is precision loss (or extra rounding). Is this good?
>

On Win64 (and any non-x86 system) the intermediary results are calculated
with Double precision while on any other x86 target it's calculated with
Extended precision before being converted to Double for the assignment to
D. So yes, that is as expected.

Regards,
Sven

>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freepascal.org/pipermail/fpc-devel/attachments/20210730/15e0005f/attachment.htm>