On 06/11/2020 18:18, Jeff King wrote:
> On Thu, Nov 05, 2020 at 09:03:49PM +0000, Ramsay Jones wrote:
>
>> diff --git a/Documentation/Makefile b/Documentation/Makefile
>> index 652d57a1b6..5c680024eb 100644
>> --- a/Documentation/Makefile
>> +++ b/Documentation/Makefile
>> @@ -272,7 +272,9 @@ install-html: html
>> ../GIT-VERSION-FILE: FORCE
>> $(QUIET_SUBDIR0)../ $(QUIET_SUBDIR1) GIT-VERSION-FILE
>>
>> +ifneq ($(MAKECMDGOALS),clean)
>> -include ../GIT-VERSION-FILE
>> +endif
>
> Calling out "clean" here specially feels somewhat backwards to me, in
> terms of Makefile design. In an ideal world we provide all of the
> dependencies to "make", and based on the targets we are building, it
> decides what needs to be done.
>
> This works with normal targets, obviously, but also with variables. If
> we do:
>
> FOO = $(shell do-the-thing)
>
> then we execute that command only when $(FOO) is needed[1].
>
> But "include" here is tricky. It is loaded regardless of whether the
> values it contains are needed or not. I wonder if we could do better by
> giving make more information about what we're expecting to get from it.
> I.e., if we wrote:
>
> GIT_VERSION = $(shell awk '{print $3}' GIT-VERSION-FILE)
>
> Then "make clean", not needing the value of $(GIT_VERSION), wouldn't run
> that shell snippet at all. Of course there's a catch; we are also
> relying on the include to trigger the dependency. So it is really more
> like:
>
> GIT_VERSION = $(shell make GIT-VERSION-FILE && awk '{print $3}' GIT-VERSION-FILE)
>
> I'm not sure how bad that is. Re-invoking make seems like it could get
> expensive, especially for the common case that we're building actual
> binaries and we _do_ need the version. But we could probably cut "make"
> out of the loop entirely. Generating GIT-VERSION-FILE is already a FORCE
> target, so really:
>
> GIT_VERSION = $(shell ./GIT-VERSION-GEN)
>
> would be equivalent-ish (with some output changes, and possibly we'd
> want to stash the value in a file for any other scripts to make use of).
>
> This is all just stuff I've written in my editor and not tried. I won't
> be surprised if there are some gotchas. But it at least seems like a
> conceptually cleaner path.
>
> -Peff
>
> [1] Variable assignment actually has a slight problem in the opposite
> direction: it wants to run the shell snippet every time the variable
> is referenced. There's a trick to get around that described in
> 0573831950 (Makefile: avoid running curl-config unnecessarily,
> 2020-04-04).
>
> It's built around evals. In fact, I suspect you could build a
> function around eval that actually works similar to include, but
> lazy-loads the file only when one of its stubs is referenced. I.e.,
> something like:
>
> GIT_VERSION = $(eval include GIT-VERSION-FILE)
>
> would probably work (and for other includes, multiple variables
> could mention the same file; as soon as it gets included, it
> overwrites the stubs).
>
Heh, in another reply in this thread, I mentioned that I had an alternate
patch I was toying with. If I tell you it was inspired by your commit
0573831950 (Makefile: avoid running curl-config unnecessarily, 04-04-2020),
you would probably not be surprised that it looks similar to what you
outline here. I had only just started looking at this approach, but it has
some wrinkles, so I thought I would look at it after submitting this series
'because this is an easy win'! ;-)
I hadn't done so yet, but I had planned to modify the GIT-VERSION-GEN script
(with -v parameter, say) to just output the version to stdout, because it
would save a sed invocation. It currently looks something like this:
diff --git a/Makefile b/Makefile
index 372139f1f2..f166fbe067 100644
--- a/Makefile
+++ b/Makefile
@@ -493,7 +493,11 @@ all::
GIT-VERSION-FILE: FORCE
@$(SHELL_PATH) ./GIT-VERSION-GEN
--include GIT-VERSION-FILE
+
+ifeq ($(wildcard GIT-VERSION-FILE),)
+$(shell ./GIT-VERSION-GEN)
+endif
+GIT_VERSION = $(eval GIT_VERSION := $$(shell cat GIT-VERSION-FILE | sed -e 's/^GIT_VERSION = //'))$(GIT_VERSION)
# Set our default configuration.
#
Ignore the 'ifeq' for now (I had several versions, including getting rid
of the GIT-VERSION-FILE rule, but that caused problems without changing
the Documentation/Makefile, and several others ... (including in contrib)).
So, I haven't worked everything out yet, but I had planned to look at
this next.
ATB,
Ramsay Jones
