[llvm-dev] ThinLTO promotion is ending up with "invalid" IR before IR-Linking

Tue Jan 3 08:29:44 PST 2017

Hi Mehdi

We observed this verification failure after our (downstream) 'mini-linker' links together two modules, not from the main llvm linker. This mini-linker is at our front end and the newly generated module is passed to the llvm optimizer.

In our downstream case, we ‘know’ about the characteristics of the modules generated in our front-end, so we can enforce the guarantee “A compile unit has at least one child” at the mini-linker phase. This avoids the verification failure around the named DICompileUnit metadata. I thought about avoiding verification of the debug info metadata (verifyModule has a flag for this), until a point where we can enforce correct debug metadata. However, correct complete verification at various phases is important, so this partial verification idea was the last resort.

Thanks,
Anna

On Dec 23, 2016, at 3:54 PM, Mehdi Amini <mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>> wrote:

Hi Anna,

Can you clarify if you observed this post-linking?

—
Mehdi

On Dec 23, 2016, at 12:42 PM, Anna Thomas <anna at azul.com<mailto:anna at azul.com>> wrote:

We had the same problem with verification when merging/linking 2 modules: a compile unit is not listed in named metadata (llvm.dbg.cu),

We worked around this problem by using the DebugInfoFinder processModule function, which identifies all the compile units in the new module, by visiting all the subprogram scopes and the already available compile units. However, this succeeds verification only if the unnamed/missing compile units have children sub programs. This is not a strong guarantee.

Anna

On Dec 23, 2016, at 2:33 PM, Mehdi Amini via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

I just hit another issue: before of ODR type uniquing on bitcode loading, we can end-up with a metadata graph that include metadata from the destination module, including a DICompileUnit.
Unfortunately this compile-unit is not listed in llvm.dbg.cu and this does not pass the verifier.

CC a few folks more aware of debug info that I am: is there a plan to make llvm.dbg.cu disappear?

—
Mehdi

On Dec 23, 2016, at 11:04 AM, Mehdi Amini via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

On Dec 23, 2016, at 9:34 AM, Teresa Johnson <tejohnson at google.com<mailto:tejohnson at google.com>> wrote:

On Thu, Dec 22, 2016 at 8:55 PM, Mehdi Amini <mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>> wrote:
Hey,

As I’m playing with Metadata lazy-loading, I added a verifier right before running the IRLinker in FunctionImport.cpp, and it does not pass (on current trunk) in a few cases. One that I looked at ended up with aliases pointing to available_externally functions for instance.

How do these look after IRLinking (in the dest module)? I looked at the logic in renameModuleForThinLTO and all the conversions to available_externally are predicated on it being in the provided GlobalsToImport set. But in FunctionImport.cpp selectCallee() we specifically prevent importing of aliases that would result in the aliasee becoming available_externally. Presumably the resulting IRLinked dest module looks legit, otherwise we would have later verifier failures.

So the source module is:

@weakalias = weak alias void (...), bitcast (void ()* @globalfunc1 to void (...)*)
define void @globalfunc1() #0 {
entry:
  ret void
}

But we turn globalfunc1 into available_externally in renameModuleForThinLTO(), which make the alias invalid.

We don’t import the IR so the destination module is OK.
36

I’m hesitant about breaking the IR verifier like that before calling the IRLinker. The alternative I can see now would be:
- to perform any handling that requires breaking the verifier as part of the IRLinker process,
- or to perform the verifier-breaking changes that we do as a pre-process step to the IRLinker as a post-process step on the linked module.

Wouldn't this second option result in verification errors on the resulting dest module?

Not sure why? Since currently the destination module is valid, the post-process should be OK as well?

—
Mehdi
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170103/40c5183b/attachment.html>