While working on the libc/sys split and all the reviews I'm stacking under D44216, I realized I needed some automation to make sure I wasn't accidentally removing symbols. This is the result. I think we should probably commit symref files to the tree for key libraries and maybe all versioned libraries as my perusal of history suggests we lose symbols by accident with some regularity.

I'm slightly ahead of myself with this change. It works fine with a native build, but needs a bootstrapped objdump in the general case. I'l sort that out in a bit.

Are you aware of devel/abi-compliance-checker? The version in ports is rotten, https://lvc.github.io/abi-compliance-checker/ seems to claim that FreeBSD is supported.

FWIW we have sym and abi checkers in junos build, they are each in a separate makefile included by bsd.lib.mk if the appropriate knob is enabled
The sym checker does BSS and TLS symbol checks, not directly useful here, but just a thought

I'm definitely open to supporting more complete checkers, but my goal is to have something in tree and enabled by default. I worry that out of tree things lead to the maintenance burden being in the wrong place as developers commit changes that haven't been checked (at least until some future day when the project has more complete pre-commit CI).

In D44271#1010069, @brooks wrote:

I'm definitely open to supporting more complete checkers, but my goal is to have something in tree and enabled by default. I worry that out of tree things lead to the maintenance burden being in the wrong place as developers commit changes that haven't been checked (at least until some future day when the project has more complete pre-commit CI).

My reference to abi-compliance-checker is to show how would our final solution for the ABI backward compatibility check look finally. I have no objections about smaller/partial checkers integrated into the build.

With that foreword, please note that I am not much versed in our make/build system.

I think I'm going to redo this using readelf. There is some annoying differences between the versions (binutil, elftoolchain, and llvm), but the format feels more manageable and I can see a path to checking for symbols changing type and weak vs strong.

Private versioning namespace should be filtered out, I think.

Our readelf adds something after the symbol+version, like

8: 00000000000bba00    17 FUNC    GLOBAL DEFAULT   13 fabs@@FBSD_1.0 (2)

Note the (2) part. It should be removed, if not already.

Perhaps the visibility column (Vis) should also participate in the final result.

In D44271#1025583, @kib wrote:

Private versioning namespace should be filtered out, I think.

I think we want to know when that set changes even if we don't consider it part of the ABI, but I probably could be persuaded otherwise. I might want to make it optional so libraries can opt in and others opt out.

Our readelf adds something after the symbol+version, like

8: 00000000000bba00    17 FUNC    GLOBAL DEFAULT   13 fabs@@FBSD_1.0 (2)

Note the (2) part. It should be removed, if not already.

awk parsing takes care of that.

Perhaps the visibility column (Vis) should also participate in the final result.

In theory I think so, but I've not found an example were it's anything other than DEFAULT so I'm not sure it adds value.

In D44271#1026288, @brooks wrote:

In D44271#1025583, @kib wrote:

Private versioning namespace should be filtered out, I think.

I think we want to know when that set changes even if we don't consider it part of the ABI, but I probably could be persuaded otherwise. I might want to make it optional so libraries can opt in and others opt out.

At least it should be split into two parts somehow, either textually in the same file, or putting the FBSDprivate into aux file at all.

In D44271#1026327, @kib wrote:

In D44271#1026288, @brooks wrote:

In D44271#1025583, @kib wrote:

Private versioning namespace should be filtered out, I think.

I think we want to know when that set changes even if we don't consider it part of the ABI, but I probably could be persuaded otherwise. I might want to make it optional so libraries can opt in and others opt out.

At least it should be split into two parts somehow, either textually in the same file, or putting the FBSDprivate into aux file at all.

After further thought, enabling --no-undefined-version is probably enough to keep track of FBSDprivate so I'll remove them from the symref file.

I noticed we mostly don't use printf in share/mk so switched to the ${:U@} trick. When doing that I noticed that bsd.symver.mk finds it's awk script with a path search so I switched to that model.