[lkml]   [2008]   [Jul]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: __weak vs ifdef

On Fri, 25 Jul 2008, Matthew Wilcox wrote:

> On Fri, Jul 25, 2008 at 02:34:55AM -0700, Andrew Morton wrote:
> > We should make arch_pick_mmap_layout __weak and nuke that ifdef.
> I strongly disagree. I find it makes it harder to follow code flow
> when __weak functions are involved. Ifdefs are ugly, no question, but
> they're easier to grep for

Hell no, they're not.

Our use of random HAVE_ARCH_xyz or ARCH_SUPPORTS_xyz etc stuff makes
things _totally_ impossible to grep for.

In contrast, it we did this code as

#ifndef arch_pick_mmap_layout
void __weak arch_pick_mmap_layout(struct mm_struct *mm)
mm->mmap_base = TASK_UNMAPPED_BASE;
mm->get_unmapped_area = arch_get_unmapped_area;
mm->unmap_area = arch_unmap_area;

then trying to grep for arch_pick_mmap_layout() would show EVERY SINGLE
RELEVANT CASE! And it would show the "__weak" there too, so that once
people get used to this convention, they'd have a really easy time
figuring out the rules from just the output of the 'grep'.

I really think that whoever started that 'HAVE_ARCH_x'/'ARCH_HAS_x' mess
with totally random symbols that have NOTHING WHAT-SO-EVER to do with the
actual symbols in question (so they do _not_ show up in grep'ing for some
use) should be shot.

We should never _ever_ use that model. And we use it way too much.

We should generally strive for the simpler and much more obvious

/* Generic definition */
#ifndef symbol
int symbol(..)

and then architecture code can do

#define symbol(x) ...

or if they want to do a function, and you _really_ don't like the '__weak'
part (or you want to make it an inline function and don't want the clash
with the real declaration), then you can just do

static inline int symbol(x)
#define symbol symbol

and again it all works fine WITHOUT having to introduce some idiotic new
and unrelated element called ARCH_HAS_SYMBOL.

And now when you do 'git grep symbol' you really will see the rules. ALL
the rules. Not some random collection of uses that don't actually explain
why there are five different definitions of the same thing and then you
have to figure out which one gets used.

> My basic point here is that __weak makes the code easier to write but
> harder to read, and we're supposed to be optimising for easier to read.

But your basic point is flawed. The thing you advocate is actually harder
to read.

Yes, if you don't follow the codign style, and you write

int __weak

you are (a) a moronic rebel you never understood why the declaration
should be on one line and (b) as a result your 'grep' won't see the __weak
and you'll be confused about the rules.

But if we _consistently_ used

- '#ifndef symbol' to avoid redeclaring something that the architecture

- and '__weak' to allow architectures to just override functions without
extra work and rules

then after a while people would simply _know_ that very simple set of
rules, and a 'grep' would work so much better than it does now.

Really. Try it. Try it with 'arch_pick_mmap_layout' (with Andrews patch in
place). And then imagine that you'd be used to '__weak', and seeing that

mm/util.c:#ifndef arch_pick_mmap_layout
mm/util.c:void __weak arch_pick_mmap_layout(struct mm_struct *mm)

in the output. Be honest now - wouldn't that actually _tell_ you something
relevant about that particular declaration? And make the fact that some
architectures override it _less_ confusing?

IOW, you could tell directly from the grep output that it's a "default
fallback". Which you definitely cannot tell right now, because we have
insane models for doing it.


 \ /
  Last update: 2008-07-26 21:43    [W:0.103 / U:0.432 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site