Alex Rousskov wrote:
On Mon, 2008-03-10 at 18:59 +1300, Amos Jeffries wrote:
Alex Rousskov wrote:
IMO, at the end of the cleanup, "squid.h" should be the file that all
sources are guaranteed to include first simply because it is called
squid.h. Config.h has, by name, a much narrower scope.
I realize that this is subjective and that very few people care about
things like that in Squid so if, for whatever reason, Amos wants to
remove squid.h from most files and leave config.h instead, I am not
going to object since I am not the person doing the cleanup leg work,
and since the functional effect will be the same.
The little tweaks that are not global in nature should be moved out to
specific headers and wrappers, of course. The
guaranteed-to-be-included-first.h file should eventually be almost
empty. Its primary value is the guarantee itself, which becomes a useful
tool for [usually temporary] hacks.
With the testheaders.sh every .h now has
guaranteed-to-be-included-at-earliest-actual-need.
Yes, that is great(*), but to-be-included-first guarantee is a little
different (and much easier to support/test).
The hacks should be localized to the object(s) they are hacking anyway
and not affect the entire Squid + sub-programs + linked-programs +
helpers + unit-tests + tools.
If they do you would be:
- hacking a system include
(which requires config.h for the wrapping!)
- hacking an objects behaviour
(which gets linked/included everywhere its used anyway!)
The guaranteed-to-be-included-first.h file is needed for exceptional
hacks that defy "normal" rules. For example, it is 3:00 in the morning
and you must make Squid to compile on some remote box, but some internal
system header is broken and needs a special _BSD_API #define to compile.
Many system headers include that header conditionally so you cannot just
quickly wrap a system include file to #define that constant. If only you
could define it before any system header is included!
Such hacks are usually temporary.
Those hacks would go in config.h around #include "autoconf.h". Which in
squid is guaranteed-before-system-headers.
The few places where the sys headers are not wrapped need fixing.
Again, I am just sharing my experience here. No significant pressure to
implement guaranteed-to-be-included-first.h file idea...
Thank you,
Alex.
P.S. Can a script reliably test an included-at-earliest-actual-need
guarantee? I think I can write a header that compiles fine without
config.h but still needs it to enable an optional API. How would a
script detect such a need?
Thats the tricky bit. It can test the guarantee for included files and
custom types. Which covers most of the cases, where the optional API is
represented by a class or uses new branch-specific fields of objects.
What can't be tested this way is the inverse case of code which does not
add object fields anywhere, reference branch-specific fields, or use a
branch-specific class API. That case can only really be tested by hand.
Amos
--
Please use Squid 2.6STABLE17+ or 3.0STABLE1+
There are serious security advisories out on all earlier releases.
