[Lldb-commits] [PATCH] D58564: [Reproducers] Add command provider
Pavel Labath via Phabricator via lldb-commits
lldb-commits at lists.llvm.org
Tue Feb 26 12:52:48 PST 2019
labath added a comment.
In D58564#1410729 <https://reviews.llvm.org/D58564#1410729>, @JDevlieghere wrote:
> In D58564#1410213 <https://reviews.llvm.org/D58564#1410213>, @labath wrote:
> > I am sorry that I won't have much time to review this in the next couple of weeks, but I don't think this is a good direction here. I don't see how this will interact with the SB API recorder, specifically with things like SBCommandInterpreter::HandleCommand, and ::HandleCommandsFromFile. The thing I would expect to see is that SB recorder captures the input of those commands (for a somewhat broad interpretation of "capture") during recording, and then substitute this during replay. That way the CommandInterpreter class would not need (almost?) any modifications.
> It’s been a while but wasn’t this exactly what you proposed in the other differential? How would you capture commands that are entered interactively (through RunCommqndInterpreter)?
> Anyway, I don’t believe this is a concern. The provider here only capture what’s entered interactively, hence the flag. Replaying the API call should work exactly as expected. I’ll double check later today.
I have to admit I haven't looked at this in detail, but the thing I'm missing here is the connection between commands and API calls. If we take RunCommandInterpreter, for instance, you can see that the lldb driver invokes this function three times. How do you ensure the "right" commands get replayed as a part of the API call?
If you look at the driver more closely, you'll see that each call to RunCommandInterpreter is preceeded by a call to SetInputFileHandle. I don't have this idea fully baked, but the way I'd try to approach this is to have each SetInputFileHandle create a new buffer where the commands will be stored in. Then as the commands are being processed (in RunCommandInterpreter), they would be added into this buffer. Then, when replaying you would know that you only should replay the commands from the given buffer.
I'd also probably try to capture these commands at a slightly lower level, because I am hoping that this will allow us to get rid of the add_to_reproducer flag. Ideally, this should fall out naturally due to the different source the commands are coming from -- the commands executed through the `HandleCommand` API would be captured at the SB boundary, and the "interactive" commands would be captured by whoever invokes the command interpreter in interactive mode.
(I have no idea how easy it is to achieve this, but that's how I'd approach this.)
CHANGES SINCE LAST ACTION
More information about the lldb-commits