[PATCH] Debug Info Verifier pass

[Manman Ren]

On Mon, Apr 14, 2014 at 5:56 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

>
> On Apr 11, 2014, at 11:31 AM, Manman Ren <manman.ren at gmail.com> wrote:
>
> > Hi Duncan,
> >
> > Thanks for working on this. The patches look great overall.
> > 0001: LGTM
> > 0002: There is some inconsistency in creating debug info verifier pass.
> > 1>
> > tools/opt/opt.cpp:  if (VerifyEach) PM.add(createVerifierPass());
> > Should VerifyEach guard debug info verifier as well given that we are
> using NoVerify to guard debug info verifier?
> > I am okay with not using VerifyEach for debug info verifier, but please
> add comments there to explain why.
>
> I just changed it to line up with Verifier for now.  Probably it will
> be too slow, but if it is, we'll probably want to add a separate flag
> so we can --verify-di-each on demand.
>
> > 2> What about these two places? Should we add debug info verifier?
> > tools/clang/lib/CodeGen/BackendUtil.cpp:
>  FPM->add(createVerifierPass());
> > unittests/Bitcode/BitReaderTest.cpp:  passes.add(createVerifierPass());
>
> I just missed those.  Thanks!  Note that one of those is in cfe, so
> I've added a third patch for that one.
>
> > 3> Could you turn debug info verifier on by default and make sure it is
> working correctly with "make check-all"?
>
> I'm seeing 225 failures with make check-all when I turn --verify-di
> and --verify-each on by default (configured with X86;ARM;ARM64):
>
>   Expected Passes    : 14993
>   Expected Failures  : 58
>   Unsupported Tests  : 2341
>   Unexpected Failures: 225
>
> make check gives:
>
>   Expected Passes    : 7893
>   Expected Failures  : 39
>   Unsupported Tests  : 2288
>   Unexpected Failures: 225
>
> so they're all in LLVM proper.
>
> Is this expected?  Given that I know next to nothing about debug
> info, I'm probably not the right person to diagnose them.
>

These tests may fail as well with the current debug info verifier when we
turn it on by default (without your patch).


>
> Should they be fixed before commit?  I'm thinking it would be
> better to tackle them incrementally.
>

I am okay with committing your patches as long as we see the same set of
failures without your patch.
We should file a bug to track these failing tests though.

Thanks,
Manman


> (For clarity, in case someone else on the list isn't following
> closely, check-all is clean unless I *force* these options on.)
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140414/52b96f13/attachment.html>