Re: p = kmalloc(sizeof(*p), )

[Al Viro <viro@ftp.linux.org.uk>]

On Sun, Sep 18, 2005 at 10:34:16PM +0200, Roman Zippel wrote:
> > Ewwwww...  I'd say that it qualifies as one hell of a bug (and yes, at least
> > 3.3 and 4.0.1 are still doing that).  What a mess...
> 
> It's not a bug, it's exactly what you're asking for, e.g. "*p1 = *p2" 
> translates to memcpy. gcc also can't simply initialize that structure in 
> place, e.g. you could do something like this (not necessarily useful but 
> still valid): "*p = (struct foo){...,  bar(p),...};".
> In the end it all depends on how good gcc can optimize away the memcpy, 
> but initially there is always a memcpy.

No.  Assignment is _not_ defined via memcpy(); it's a primitive that could
be implemented that way.  Choosing such (pretty much worst-case) implementation
in every case is a major bug in code generator.

You _can_ introduce a new local variable for each arithmetic operation in
your function and store result of operation in the corresponding variable.
As the matter of fact, this is a fairly common intermediate form.  However,
if compiler ends up leaving all these suckers intact in the final code,
it has a serious problem.

Compound literal _is_ an object, all right.  However, decision to allocate
storage for given object is up to compiler and it's hardly something unusual.
"Value of right-hand side is not needed to finish calculating left-hand side,
so its storage is fair game from that point on" is absolutely normal.