[LLVMdev] Metadata/Value split has landed
Duncan P. N. Exon Smith
dexonsmith at apple.com
Thu Dec 18 10:00:15 PST 2014
> On 2014-Dec-18, at 01:43, Keno Fischer <kfischer at college.harvard.edu> wrote:
> Hi Duncan,
> I'm in the following situation for which this change caused an assertion failure:
> Three modules, let's say A B C
> Two function, F in A, G in B
> Now I CloneFunction F into B (call the new function F') and inline F' into G.
> Now, for the problematic part, where I try to extract G (and all referenced values) into C:
> upon encountering any debug node in the inlined code, it tries to clone the DISubprogram for F', so it creates a temporary. Since that refers to F', it'll now go ahead and copy F'. However, here once again it tries to copy the DISubprogram, which now just uses the temporary value from above (this is fine). Unfortunately, right after, it calls resolveCycles on the debug info annotation, which crashes with
> Assertion failed: (!isa<MDNodeFwdDecl>(Op) && "Expected all forward declarations to be resolved"), function resolveCycles, file /Users/kfischer/julia/deps/llvm-svn/lib/IR/Metadata.cpp, line 459.
> because we still have the temporary DISubprogram in there.
> Any ideas what to do about this?
The logic I have between `MapValue(Metadata*)` and `MapValueImpl(Metadata*)`
was supposed to solve this. (Actually, those names are awful, I may try to
change them to `MapMetadata()`.)
The logical flow should be:
- MapValue(Metadata*) calls MapValueImpl().
- MapValueImpl() introduces a temporary.
- MapValueImpl() recursively calls MapValueImpl().
- MapValueImpl() resolves its temporary.
- MapValue(Metadata*) calls resolveCycles().
There could be a bug here, but I just scanned the code and it seems to
match up. I suspect, rather, that there's a temporary node in the metadata
graph on entry to `CloneFunction()`; this isn't supported, and would cause
the same assertion to fire.
Try running the verifier right before calling `CloneFunction()`: I put a
check for this in `Verifier::visitMDNode()`.
More information about the llvm-dev