<div dir="ltr"><div class="gmail_quote"><div dir="ltr">Hi all!<div><br></div><div>Fuzzing llvm's bitcode reader, I found a problem where the reader assumes that the first field in an abbreviation will not be an array or a blob (and asserts otherwise).</div><div><br></div><div>I don't know if this is expected (but not documented) or not. The documentation, to me, reads like it doesn't disallow it, but we might be assuming all abreviations start with a full record, which would make the first operand never be an array or a blob.</div><div><br></div><div>The bug comes from r181639 (<a href="http://llvm.org/klaus/llvm/commit/1197e38f3338b8db76f0fa38c2687c65b2bcea5c/" target="_blank">http://llvm.org/klaus/llvm/commit/1197e38f3338b8db76f0fa38c2687c65b2bcea5c/</a>), which took the code to read the first argument and put it outside of the loop, but didn't take the Array/Blob verification + reading code too (It's a bug because that commit was supposed to not have changed functionality :-) ).</div><div><br></div><div>This could be “fixed” with, either a report_fatal_error (if we eventually have better error handling on that code, we can make that non-fatal and report to the caller), or by hoisting the Array/Blob reading code out of the loop too (actually, write a helper function).</div><div><br></div><div>What should be done about this?</div><div><br></div><div>Thanks,</div><div><br clear="all"><div><div> Filipe<br></div></div>
</div></div>
</div><br></div>