Some background: our team has added Object to the LINK_COMPONENTS of the ProfileData library as part of work to support code coverage on baremetal emebdded devices.
Specifically, there is a new type of InstrProfReader that opens an executable and extracts unallocated sections as metadata, rather than relying on a heavyweight runtime call and metadata sections living in restricted target memory.
Recently, commit ee7d20e8 (https://reviews.llvm.org/D106861) added a dependency on the ProfileData library to the MC library.
The combination of these changes causes a dependency circuit that results in link-time failures:
- MC depends on ProfileData
- ProfileData depends on Object
- Object depends on MC
Is there any good way to resolve such dependency circuits in the build system save for an invasive restructuring?
I’ll note that in the review above, MCPseudoProbe only relies upon the static member function FunctionSamples::getCanonicalFnName, which itself relies on a number of static data members of FunctionSamples.
+firstname.lastname@example.org to bring them into the thread
That kind of static function could probably be brought out into a different library, but I think for this some sort of agreement on the dependency chains would be really helpful with what each library is bringing to the table here. It might make more sense for the MC work that’s been going on here to be happening in ProfileData with that depending on Object and MC to get work done.
Thanks for the heads-up. The MC work in discussion here deals with metadata on the binary, while ProfileData mainly deals with a profile without seeing a binary. Thus I think it would be more appropriate for the MC work to remain in MC or a third library. The name processing, i.e, calling
FunctionSamples::getCanonicalFnName can be moved into downstream consumer as a post processing. On the other hand, I’m giving a thought of completely removing the use of that function, which, if possible, should be the simplest solution.
+@Wenlei He @Lei Wang
I agree with your sentiment that the ProfileData library deals mostly with a profile and not a binary.
On my end, my team can investigate what we can do to avoid the dependency on Object from ProfileData.
Upstream, the incoming data to the reader is just a buffer of bytes. It was our downstream development that added constructs that dealt with ObjectFiles. If clients of InstrProfReader could perform the section extraction themselves and then send the profile name and data in as either StringRefs (SectionRef::getContents) or as a MemoryBuffer (getMemBuffer), we can maintain the independence of ProfData from Object. This would, however, introduce a dependence on Object for clients. Currently it looks like the only clients our development cares about is llvm-profdata, which already depends on Object, so nothing would change.
In any case, I’d appreciate being kept in the loop development regarding the removal of the call to getCanonicalFnName, in case that gains traction.
Thanks for the insightful discussion,