>
> --[[smcv]]
->> [[!template id=gitbranch author="[[smcv]]" branch=smcv/definetemplate]]
->> [[!tag patch]]
>> OK, here is a branch implementing what I said. It adds the `definetemplate`
>> directive to [[plugins/goodstuff]] as its last commit.
>>
>>>>> the right thing to do in this situation.
>>>>>
>>>>> --[[chrysn]]
+
+----
+
+[[!template id=gitbranch author="[[smcv]]" branch=smcv/ready/templatebody
+ browse=http://git.pseudorandom.co.uk/smcv/ikiwiki.git/shortlog/refs/heads/ready/templatebody]]
+[[!tag patch users/smcv/ready]]
+Branch and directive renamed to `ready/templatebody` as chrysn suggested.
+It's on-by-default now (or will be if that branch is merged).
+Joey, any chance you could review this?
+
+There is one known buglet: `template_syntax.t` asserts that the entire
+file is a valid HTML::Template, whereas it would ideally be doing the
+same logic as IkiWiki itself. I don't think that's serious. --[[smcv]]
+
+> Looking over this, I notice it adds a hash containing all scanned
+> files. This seems to me to be potentially a scalability problem on
+> rebuild of a site with many pages. Ikiwiki already keeps a lot
+> of info in memory, and this adds to it, for what is a fairly
+> minor reason. It seems to me there should be a way to avoid this. --[[Joey]]
+
+>> Maybe. Are plugins expected to cope with scanning the same
+>> page more than once? If so, it's just a tradeoff between
+>> "spend more time scanning the template repeatedly" and
+>> "spend more memory on avoiding it", and it would be OK to
+>> omit that, or reduce it to a set of scanned *templates*
+>> (in practice that would mean scanning each template twice
+>> in a rebuild). --s
+>>> [Commit f7303db5](http://source.ikiwiki.branchable.com/?p=source.git;a=commitdiff;h=f7303db5)
+>>> suggests that scanning the same page more than once is problematic,
+>>> so that solution is probably not going to work.
+>>>
+>>> The best idea I've come up with so far is to track whether
+>>> we're in the scan or render phase. If we're in the scan
+>>> phase, I think we do need to keep track of which pages
+>>> we've scanned, so we don't do them again? (Or perhaps that's
+>>> unnecessary - commit f7303db5 removed a scan call that's in
+>>> the render phase.) If we're in the render phase, we can assume
+>>> that all changed pages have been scanned already, so we can
+>>> drop the contents of `%scanned` and rely on a single boolean
+>>> flag instead.
+>>>
+>>> `%scanned` is likely to be no larger than `%rendered`, which
+>>> we already track, and whose useful lifetime does not overlap
+>>> with `%scanned` now. I was tempted to merge them both and call
+>>> the result `%done_in_this_phase`, but that would lead to really
+>>> confusing situations if a bug led to `render` being called sooner
+>>> than it ought to be.
+>>>
+>>> My ulterior motive here is that I would like to formalize
+>>> the existence of different phases of wiki processing - at the
+>>> moment there are at least two phases, namely "it's too soon to
+>>> match pagespecs reliably" and "everything has been scanned,
+>>> you may use pagespecs now", but those phases don't have names,
+>>> so [[plugins/write]] doesn't describe them.
+>>>
+>>> I'm also considering adding warnings
+>>> if people try to match a pagespec before scanning has finished,
+>>> which can't possibly guarantee the right result, as discussed in
+>>> [[conditional_preprocess_during_scan]]. My `wip-too-soon` branch
+>>> is a start towards that; the docwiki builds successfully, but
+>>> the tests that use IkiWiki internals also need updating to
+>>> set `$phase = PHASE_RENDER` before they start preprocessing. --s
+
+>>>> reviewing those modifications, i think this is a good way to go. along
+>>>> with warning about pagespecs evaluated in scan phase, i think it should be
+>>>> an error to invoke scan in the render phase; that would mean that
+>>>> `readtemplate` needs to check whether it's invoked as a scan or not to
+>>>> decide whether to scan the template page, but would be generally more
+>>>> robust for future plugin writing. --[[chrysn]]