Subject: [Boost-build] detection of dependency changes during a build
From: Ray Lambert (codemonkey_at_[hidden])
Date: 2010-03-20 17:10:01
I'm attaching a small test case that I put together to demonstrate an
BB that has been preventing me from accomplishing certain goals. I won't go
into what those goals are at the moment as I just want to discuss this
particular issue. Note that some aspects of the test case are contrived in
order to achieve the desired demonstration.
What I perceive to be happening is that BB reads and caches timestamps for
dependencies early (in the parsing phase, perhaps) and therefore does
changes that are made to those files later in the build process (in the case
where a tool that it invokes on behalf of the user changes a dependency).
The elements of the test case are:
- foo.h defines a symbol as such: #define FOO 1
- foo.cpp includes foo.h and prints FOO
- foo is built via Jamroot as such: exe foo : foo.cpp ;
- foo.sh is a shell script that reads the number in foo.h,
and then re-writes foo.h with the new number
- foo.sh is invoked via Jamroot using the SHELL rule
There's also a script called test.sh which runs the test. It initializes by
re-writing foo.h to define the value "1" and deleting the bin directory. It
then alternately invokes bjam and foo several times while displaying the
contents of foo.h.
What actually happens is this:
On the first build, FOO is initially "1" and foo and foo.o don't exist.
incremented to "2" and foo is built. foo correctly displays "2".
On the second build, FOO is incremented to "3" but foo is *not* built
(apparently because foo.h's *intial* timestamp did not indicate that
was out of date). foo now *incorrectly* displays "2".
On the third build, FOO is incrementeed to "4" and foo *is* built. This
the initial foo.h timestamp showed that foo.o was out of date. foo now
correctly displays "4".
On subsequent builds, FOO is incremented every time but foo is only rebuilt
every other time. As a result, foo will never display an odd number.
The obvious questions here are:
- Am I doing something wrong?
- Is there a way to work around this?
- Is this a bug (or, more likely, a design issue)?
- Is this something that was inherited from jam and is therefore
impossible to fix?
Opinions may differ on this but I consider this to be incorrect behavior. I
would expect a build tool to detect changes made to dependencies during
Theoretically (and assuming my suspected cause is correct), this can be
changing BB to wait until the last possible moment to check dependency
timestamps. I realize, though, that may be difficult or impossible in
Any comments are welcome.
-- In a world without walls and fences, who needs windows and gates?
Boost-Build list run by bdawes at acm.org, david.abrahams at rcn.com, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk