svn commit: r262696 - head/sys/arm/arm

[Bruce Evans]

On Sun, 2 Mar 2014, Ian Lepore wrote:

> Log:
>  Add __attribute__((used)) so that the delay implementation doesn't get
>  optimized away as unreferenced, causing linker errors when trying to
>  resolve the weak reference to the missing function.

Why not use the standard FreeBSD macro __used?  Hard-coded
__attribute__(())s are not only unportable; they are gross style bugs.
They were all replaced by FreeBSD macros in old versions of FreeBSD.

> Modified: head/sys/arm/arm/mpcore_timer.c
> ==============================================================================
> --- head/sys/arm/arm/mpcore_timer.c	Sun Mar  2 19:46:03 2014	(r262695)
> +++ head/sys/arm/arm/mpcore_timer.c	Sun Mar  2 21:25:32 2014	(r262696)
> @@ -370,6 +370,7 @@ DRIVER_MODULE(mp_tmr, simplebus, arm_tmr
>  *	nothing
>  */
> static void
> +__attribute__((used)) /* Must emit function code for the weak ref below. */
> arm_tmr_DELAY(int usec)
> {
> 	int32_t counts_per_usec;

The bug was really in the weak reference macro.  It is hard-coded in the
same unportable asm for all arches and compilers (so the unportable gccism's
in it are much larger than for __attribute__(())).  Thus the compiler cannot
see it.  It is missing C parts that declare the referred-to identifier as
__used.

The __weak_reference() macro has dubious support for __STDC__ (it doesn't
really suoport !__STDC__, but only cpp -traditional).

Some macros in sys/cdefs.h has been broken to pretend to support lint.
This actually breaks lint, by defining away the macros so that lint
can't detect their unportabilities.  One of these is __unused.  But the
support for lint is so broken that __unused is not one of these.  Neither
is __weak_reference() or __attribute__().  The last is intentional.  All
of the simpler macros like __used and __unused are just aliases for
__attribute__(()).  These should be defined away for lint if and only if
they are just optimization hints or other hints not related to the
correctness of the program, e.g., __pure2.  The others, and __attribute__()
itself, should be left undefined so that lint barfs on them if they are
are ever used.

lint has a -g flag that supports some gcc extensions.  These don't include
__attribute__(), so -g was almost useless in practice in 1993.  It is more
useless now.  lint even has a hint about unconditionally defining away
__attribute__().  It has a line of source to do this, but this line is
ifdefed out in FreeBSD.

Bruce