Re: [Qemu-devel] [PATCH] const / static (against current CVS)

[Andreas Mohr]

Hi,

On Wed, Aug 24, 2005 at 02:41:44PM +0100, Paul Brook wrote:
> > Probably more important is to make sure none constant data structures
> > are done on the stack.  There is no good reason why any code page
> > should be read-write.
> 
> Huh? this is nonsense.

Uhoh, I seem to have managed to stir up a veritable hornet's net with my
patch...

> Making global data readonly is a small win because it means it can be shared 
> by multiple instances of the same application.

Yup.

> Moving global RW data onto the stack isn't neccly a win. Most systems have a 
> relatively small limit on stack size, so putting large objects on the stack 
> is a bad idea.

Indeed. Witness e.g. Linux's optional kernel stack reduction to 4k
and the problems that ensued.

> Contrary to popular belief the "const" qualifier on pointers has absolutely 
> no 
> effect on optimization. It's simply a debugging aid so the compiler will 
> generate an error if you accidentally assign to it. It's perfectly legal to 
> cast a (const char *) to a (char *) then dereference and write to it, 
> provided the object the object it points to is modifiable.

I was thinking just that: that constifying stuff at most gives a very small
compiler gain, but most probably just nothing (some people reported some
rather weak 1.4% performance gain somewhere!? Statistical noise...).
Since you say it's nothing I'm strongly inclined to believe it...

Well, it's not only a compiler debugging aid, but also a runtime debugging
aid, since random (attempted) .rodata page corruption should SEGV in that case.

Andreas Mohr