Hi Everyone, In September I sent out an RFC [1] about adding reproducers to LLDB. Over the past few months, I landed the reproducer framework, support for the GDB remote protocol and a bunch of preparatory changes. There's still an open code review [2] for dealing with files, but that one is currently blocked by a change to the VFS in LLVM [3].
The next big piece of work is supporting user commands (e.g. in the driver) and SB API calls. Originally I expected these two things to be separate, but Pavel made a good case [4] that they're actually very similar. I created a prototype of how I envision this to work. As usual, we can differentiate between capture and replay. ## SB API Capture When capturing a reproducer, every SB function/method is instrumented using a macro at function entry. The added code tracks the function identifier (currently we use its name with __PRETTY_FUNCTION__) and its arguments. It also tracks when a function crosses the boundary between internal and external use. For example, when someone (be it the driver, the python binding or the RPC server) call SBFoo, and in its implementation SBFoo calls SBBar, we don't need to record SBBar. When invoking SBFoo during replay, it will itself call SBBar. When a boundary is crossed, the function name and arguments are serialized to a file. This is trivial for basic types. For objects, we maintain a table that maps pointer values to indices and serialize the index. To keep our table consistent, we also need to track return for functions that return an object by value. We have a separate macro that wraps the returned object. The index is sufficient because every object that is passed to a function has crossed the boundary and hence was recorded. During replay (see below) we map the index to an address again which ensures consistency. ## SB API Replay To replay the SB function calls we need a way to invoke the corresponding function from its serialized identifier. For every SB function, there's a counterpart that deserializes its arguments and invokes the function. These functions are added to the map and are called by the replay logic. Replaying is just a matter looping over the function identifiers in the serialized file, dispatching the right deserialization function, until no more data is available. The deserialization function for constructors or functions that return by value contains additional logic for dealing with the aforementioned indices. The resulting objects are added to a table (similar to the one described earlier) that maps indices to pointers. Whenever an object is passed as an argument, the index is used to get the actual object from the table. ## Tool Even when using macros, adding the necessary capturing and replay code is tedious and scales poorly. For the prototype, we did this by hand, but we propose a new clang-based tool to streamline the process. For the capture code, the tool would validate that the macro matches the function signature, suggesting a fixit if the macros are incorrect or missing. Compared to generating the macros altogether, it has the advantage that we don't have "configured" files that are harder to debug (without faking line numbers etc). The deserialization code would be fully generated. As shown in the prototype there are a few different cases, depending on whether we have to account for objects or not. ## Prototype Code I created a differential [5] on Phabricator with the prototype. It contains the necessary methods to re-run the gdb remote (reproducer) lit test. ## Feedback Before moving forward I'd like to get the community's input. What do you think about this approach? Do you have concerns or can we be smarter somewhere? Any feedback would be greatly appreciated! Thanks, Jonas [1] http://lists.llvm.org/pipermail/lldb-dev/2018-September/014184.html [2] https://reviews.llvm.org/D54617 [3] https://reviews.llvm.org/D54277 [4] https://reviews.llvm.org/D55582 [5] https://reviews.llvm.org/D56322
_______________________________________________ lldb-dev mailing list lldb-dev@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/lldb-dev