*** Bug llvm/llvm-bugzilla-archive#927 has been marked as a duplicate of this bug. ***

replace-annotation-with-listener.patch
This patch simply replaces the subversive and overgeneralized annotation
classes with a map in MachineFunction and a listener interface so that
MachineFunctions can destroy themselves when the parent Function is deleted.

I wish I'd noticed this one before 927. Not sure you'll want to apply this
patch in light of the comments on this bug.

Are we going to apply this patch?

Reid.

No, I'm working on a better one incorporating some feedback from Chris, but have been rather busy with
work.

mfprovider.tar.gz
This patch implements the strategy I discussed with Chris. An "analysis" pass
is used as a bucket to hold onto the MachineFunction generated by instruction
selection. PassManager then releases the memory at the appropriate time. I find
this approach is rather fragile (the caveats are noted in the header file), but
it's much less disruptive than adding direct support for MachineFunctionPasses
to PassManager et al.

Please note that the attachment is a .tar.gz to capture the new files; it
should be untarred in the root of the source tree, and the patch applied.

Did this patch ever get applied? What's the status?

We could save 4 bytes per Function and avoid multiple inheritance if we made
Function not derive from Annotable.

It would be nice to close this (finally) for 2.0

It was not applied, no.

The patch has bitrot. There are some trivial mechanical conflicts which are easily resolved. Additionally,
llvm-gcc's pipeline now invalidates the MFProvider, which trips an assert; the most obvious cause for that
would be a MachineFunctionPass failing to call super in its getAnalysisUsage override.

This is done!