On 05/18/2014 02:33 PM, Alan Mackenzie wrote:
> On Fri, May 16, 2014 at 11:06:24AM -0700, Daniel Colascione wrote:
>> On 05/16/2014 10:52 AM, Alan Mackenzie wrote:
>>>>>> Today, we use clever macros to hard-code the values of all cc-mode
>>>>>> language variables into the mode functions of each cc-mode major mode
>>>>>> function or into c-init-language-vars-for, but in order to allow users
>>>>>> to customize cc-mode syntax, we have to be able to recompute language
>>>>>> constants and variables at runtime.
> 
>>>>> Do we, now?  You can imagine I've one or two reservations about this
>>>>> idea. 
> 
>>>> What's your alternative?
> 
>>> Turning the pertinent c-lang-defvars, and only these, into configurable
>>> variables in cc-vars.el.
> 
>> Have you looked into what task actually involves?
> 
> No I hadn't.  
> 
>> You need to modify c-decl-hangon-kwds, and as a result,
>> c-prefix-spec-kwds, c-postfix-spec-kwds, c-keywords,
>> c-keywords-regexp, c-keywords-obarray, c-nonlabel-token-key, and
>> c-regular-keywords-regexp.
> 
> OK.  Then another alternative is to use `or' forms in the appropriate
> places, of which there are several rather than many, the extra arm of
> the `or' looking something like "(looking-at c-noise-macro-re)".  Did you
> consider this at all?  If so, why do you prefer the solution you now
> propose?

Doing it that way further increases the amount of work needed at parse
time, and cc-mode isn't a speed demon already. Your proposal also
introduces much-increased for complexity in the core and creates a risk
of "missing a spot" and introducing a bug where we check the primary
keywords but not the user-customized ones. (And what about the checking
we do after calling intern-soft on the symbol obarray?) cc-engine is
very difficult to follow as it is. I'd rather gracefully extend the
existing functionality than add yet another set of special cases.

Another approach I was considering was to change cc-mode to always skip
over regions marked with a certain text properly. Then a minor mode
could independently parse and fontify the buffer, marking regions it
wanted cc-mode to ignore with this property. This approach is more
flexible, but I don't think the complexity is worth it when a simple
extension to cc-mode is within reach.

>> It's easier and more flexible to simply allow the entire set of
>> c-lang-defconst values to change.
> 
> It's the flexibility that worries me.  Flexibility means complexity (of
> which there's already too much in CC Mode), and might open up a larger
> "attack surface" for future bugs.
> 
> I've spent a lot of time trying to pin down exactly what your proposed
> change does and how it does it, but amn't there yet.  I only have the
> before and after versions, not a detailed description of the change.

I'd hoped my initial message clearly explained the overall approach.
Briefly, we want to allow users to tell cc-mode to ignore "noise macros"
when parsing. IMHO, the best way to do that is to let user
customizations augment the existing set of "noise macro" keywords
cc-mode defines for each language. Today, cc-mode builds various data
structures from these keyword sets at compile time. In order for user
customizations to have any effect, we have to build these data
structures at runtime instead. Doing that has no perceivable effect on
mode initialization performance and lets user customizations have an effect.

>> You claim that there is a great risk of negative side effects resulting
>> from this change.
> 
> No.  I've said there is a risk, without quantifying it.
> 
> One question which you haven't addressed yet is does the changed code
> work properly on modes derived from a CC Mode mode?  How easy is it to
> create a `foo-extra-keywords' for such a mode?

Trivial --- just include the variable in the c-lang-defconst calculation
for the derived mode.

> Another is does the change work properly under XEmacs?

I don't see why it wouldn't.

> What is the purpose of {java,idl,pike}-extra-keywords, given that these
> three languages don't have #defines?

Consistency with the type defcustoms? Accommodations for weird custom
preprocessors? Why not?

>> I don't agree, and I don't see anyone else proposing, much less
>> implementing, a solution to this problem, .....
> 
> :-)  This problem was first highlighted as such, by you, this month, not
> several years ago.

There's been a TODO in the code to this effect since 2005 in the
definition of c-decl-hangon-kwds; my patch removes this TODO.

> Your patch is complicated, it's disruptive, yet you
> seem put out that I want to discuss it and understand it, rather than
> just blindly applying it as it is.  It seems your use of "You need to"
> and "we have to" weren't mere hyperbole as I first thought.  Forgive me,
> but I'm always suspicious when somebody tells me "there's no
> alternative".  There are always alternatives.  The question is why is
> your patch, which is not a straightforward or obvious change, the best
> solution to the problem.  What are the tradeoffs?

I discussed some of them above. The chief risk is that users might
customize cc-mode by altering its language constants. I strongly
discount this risk, especially since users can *already* do that by
creating derived modes.

>> .... and cc-mode's being developed outside the tree makes it
>> frustratingly slow and difficult to get much-needed fixes into core
>> where users can see them.
> 
> Would it really be much faster if CC Mode were integrated?  Not really.
> What makes things slow is the small number of people working on CC Mode,
> i.e. one, with occasional help from people like yourself.
> 
>>> As I said, I'm not at all happy at making such a massive change to CC
>>> Mode's architecture.  There would surely be unforeseen consequences, some
>>> of which might well be negative.
> 
>> It's not a massive change in architecture.
> 
> As I said, I don't fully understand the change, yet.  But moving
> calculations from compile time to run time _is_ a change of
> architecture, and it seems to me not to be small.

All the cc-defvar variables get the same values they did before, by
default. cc-engine parses text the same way it always has. All the
cc-mode customization remains. My patch merely lets users customize some
existing cc-mode constants, then makes these customizations have runtime
effect by removing some unnecessary old optimizations in cc-mode
initialization. If you call that a massive change in architecture, then
what I call architecture might as well be plate tectonics.

> It seems to me
> that the rigid separation between language (e.g. C) and user
> configuration is being broken down.  How much easier will it be for a
> user mistakenly to attempt to configure her CC Mode by messing around
> with the c-lang-consts?

Much easier --- but it's never been a priority in Emacs to prevent users
from hanging themselves. Users should be drawn to the customization
settings we provide, and if they choose to bypass them, they probably
know well enough what they're doing. As I mentioned, determined users
can already derived from cc-mode built-in modes.

>> cc-mode can already evaluate these variables at runtime in the case of
>> a version mismatch.
> 
> That's a feature intended purely for CC Mode developers.

Or for people who upgrade without recompiling derived modes. But it
doesn't matter --- the dynamic-computation code is there and works fine.
My change just uses it all the time instead of going down a different
path involving dubious compile-time optimizations.