A recent change to the way metadata works has broken my StaticRoots pass,
and I need some advice.
I was using metadata to store information about static roots because of a
convenient feature of metadata nodes: the arguments to an MDNode are treated
as weak references, which means that a metadata node that points to a global
value won't prevent the dead global pass from removing that global. This is
very handy for managing static garbage collection roots, because ordinarily
the static roots table contains pointers to the global variables, and those
pointers would prevent the global from being considered dead - which means
that your final executable would have a bunch of unused cruft in it. But by
storing the static roots in metadata, and then constructing the root table
from that metadata after the global elimination pass has finished, we end up
with a roots table that only includes the roots that are actually used.
So my first question is - am I abusing the metadata system by doing this?
That is, am I relying on undefined behavior here? The reason I ask this is
because the scheme I outlined in the previous paragraph no longer works with
the current LLVM tip - the metadata nodes that used to contain references to
global variables are now all NULL, regardless of whether the global variable
was dead or not.
My second question is, if metadata is not the right way to do this, then
what is? An alternate approach would be to have two passes, one which runs
before the dead global pass, and one which runs after. The first pass
removes the static roots table from the module (so that it won't cause dead
globals to become live), and the second pass adds it back, minus the global
vars that were removed. The problem is guaranteeing that the passes are
executed in the right order - that is, I want pass A to execute before the
dead global pass, and pass B to execute after it. Is there any way I can get
the pass manager to schedule my passes like this? I would prefer not to have
to manually specify the order of passes on the command line, I'd rather just
have the pass manager do it's work.
--
-- Talin
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20110805/6628be8c/attachment.html>