Re: POINTER_ALIGNED_P (was: Re: CVS commit: src/sys)

[Valery Ushakov <uwe%stderr.spb.ru@localhost>]

On Wed, Feb 17, 2021 at 00:51:03 +0100, Roland Illig wrote:

> 17.02.2021 00:25:10 Valery Ushakov <uwe%stderr.spb.ru@localhost>:
> > On Tue, Feb 16, 2021 at 23:54:46 +0100, Roland Illig wrote:
> >> Yes, it does.  That's what the "#undef __NO_STRICT_ALIGNMENT" in the
> >> test is for.
> >>
> >> I intentionally placed it between <sys/types.h> (which defines that
> >> macro on x86 and some other platforms) and <sys/param.h> (which uses the
> >> macro to switch between the boring "everything is correctly aligned" and
> >> the more interesting formula suggested here.
> >
> > This is wrong on so many levels.  What is even the point of a test
> > that doesn't test the thing as it is actually defined and used in the
> > code?
> 
> The point of the test is to verify that the "complicated" formula
> produces correct results.  That's what several commits tried to
> fix.  If this test had been there from the beginning, none of the
> wrong formulas would have passed it.  That's the whole point.
> 
> The point of the test was intentionally not to test the actual
> behavior on each platform but to test the same formula, independent
> from the platform, and to do this, I somehow needed access to that
> formula.  Testing the actually used formula per platform could be
> added as another test.  I just wanted to avoid the obviously wrong
> formulas to go unnoticed in the code.  That's the point of the test,
> and that's exactly what it achieves.  Therefore I don't see anything
> wrong with it.

The very fact that you need to undefine an unspecified macro at an
unspecified time to get to the "formula" points to a problem.  We
shouldn't be pretending that it's not, and provide the false decorum
of "oh, but it's covered with a test, so it's ok".

-uwe