Re: Deduplicating __NR_* and SYS_* syscall defines

[Rich Felker <dalias@libc.org>]

On Wed, May 11, 2016 at 07:22:30PM -0500, Bobby Bingham wrote:
> On Wed, May 11, 2016 at 05:53:12PM -0400, Rich Felker wrote:
> > On Mon, May 09, 2016 at 07:26:37PM -0500, Bobby Bingham wrote:
> > > During the powerpc64 review, Rich mentioned wanting to replace the
> > > arch/*/bits/syscall.h files with .in files that would be preprocessed with
> > > something like:
> > > 
> > >     sed -e p -e s/__NR_/SYS_/ < $< > $@
> > > 
> > > This would eliminate a lot of intra-file duplication here.
> > > 
> > > I took a look, and this won't quite work as-is, because the following
> > > lines in the arm version would end up outputting duplicate definitions:
> > > 
> > >     #define __ARM_NR_breakpoint	0x0f0001
> > >     #define __ARM_NR_cacheflush	0x0f0002
> > >     #define __ARM_NR_usr26		0x0f0003
> > >     #define __ARM_NR_usr32		0x0f0004
> > >     #define __ARM_NR_set_tls	0x0f0005
> > 
> > This is easily fixed by something like:
> > 
> > sed -e /__NR_/p -e s/__NR_/SYS_/ < $< > $@
> 
> Neat.  I didn't know about /p.

The p command just prints the pattern space. The trick is that we
print it an extra time before the s command, but only if the line
matches something the s command is going to change.

> Any objection to using Alexander's approach to avoid interleaving the
> __NR_* and SYS_* lines?

I'm indifferent to the interleaving, but if we take that approach, we
should make sure that the rules are written such that interrupting the
make process between the commands doesn't leave a partial file that
subsequent runs of make think is complete. It might be ok as-is if
make automatically deletes the target on error producing it; otherwise
we might need a temp file that's moved into place at the end. I always
forge how this aspect of make works...

> > > Same thing for this line in x32:
> > > 
> > >     #define __X32_SYSCALL_BIT        0x40000000
> > 
> > In general we've tried to eliminate this sort of macro and direct-code
> > the values. I would be in favor of doing the same for x32 I think. But
> > with my fixed sed command (above) I think that change is unnecssary
> > and orthogonal to the deduplication.
> 
> I'll submit a separate patch to clean this up.

Any thoughts on how it should be done? If it were a clean decimal
constant like on mips I'd just write each as a single integer literal
(e.g. 6001, etc.) but since the syscall numbers are normally thought
of as decimal whereas the x32 offset is hex/bit value, it seems + or |
is still needed.

> > > I'm thinking something like the following awk script would work:
> > > 
> > >     {
> > >         print
> > >     }
> > > 
> > >     $1 ~ /^#(define|undef)$/ && $2 ~ /^__NR_/ {
> > >         sub(/__NR_/, "SYS_", $2)
> > >         print
> > >     }
> > > 
> > > The handling for #undef is for the x32 file.  It looks like only the
> > > `#undef __NR_getdents' in that file is actually necessary, and even that
> > > could be avoided by just omitting the earlier line:
> > > 
> > >     #define __NR_getdents (__X32_SYSCALL_BIT + 78)
> > 
> > I don't see why any #undef is needed here; this looks like leftover
> > cruft that was not properly cleaned up. All the logic for replacing
> > syscall numbers belongs in src/internal/syscall.h or
> > arch/$ARCH/syscall_arch.h, I think.
> 
> Ok.

Can you verify that removing these #undef-and-redefine lines from the
bits header doesn't change the result (i.e. it's redundant with other
files) for building libc?

> > > So maybe we can get rid of the #undefs there, and simplify the awk script
> > > accordingly.
> > > 
> > > Thoughts on this approach?  If this sounds ok, I'll submit a patch.
> > 
> > I'd rather use sed than awk if possible since it's more universally
> > available and understood.
> 
> I must be the exception then :)

:-)

Rich