Global lldbinit file?

When LLDB launches, it will read ~/.lldbinit (and possibly other variants of that, like ~/.lldbinit-Xcode).

In my Hexagon LLDB installation, I want LLDB to always load a global lldbinit file. Currently I do this by having lldb be a wrapper script that calls lldb with –o, but I’d like to eliminate the wrapper script, since wrapper batch files cause problems with ctrl-c handling on Windows. I use this to load a python file with utilities in it, like one to get the TLB info from the target. It sends a qXfer command to the remote GDB server to download an XML file, parses it, and prints the info from it. These utilities need to load for all users.

Do we have a global lldbinit file, or should I add code to load Host::GetProgramFileSpec()/lldbinit?

This worries me a little bit. I install your Hexagon lldb, and that installs a global lldbinit, which then starts affecting my use of the other lldb's I happen to have on my system in ways that are not at all clear to me... That doesn't seem like a good idea to me.

gdb had a global configuration file in /etc/conf, and Apple's gdb installation used it for a few things (this was actually done at NeXT and NOT by me...) - mainly setting some common print variables. But this just ended up causing more confusion than it was worth. It would have been better to just bake these into the gdb driver we shipped...

I would suggest distributing your lldb as lldb-hex or something like that, and including a template .lldbinit-hex that folks could install, or making your own driver that sets the options you want first.

Jim

I'm sorry; I wasn't clear. When I say "global", I mean "global to this installation", not "global to the system".

My idea was something like /inst/hexagon/Tools/bin/lldb would load /inst/hexagon/Tools/bin/lldbinit. Another lldb installation, say /inst/Xcode/bin/lldb, wouldn't see it, or be affected by it.

A few questions:

Can you build it into your LLDB to it is hard coded? Why keep a separate text file? Can you trigger the hexagon stuff to happen once and only once, but only after a hexagon debug session is initiated? I would prefer this lazy initialization if possible.

That's not so bad, though I don't think putting config files in the bin directory is going to be a very popular thing to do. But I still feel that this ends up being a very opaque way to change the behavior of lldb. If the changes are trivial, then the benefit of doing this doesn't seem worth the potential confusion, and if they are significant, then it seems to me better to make a special-purpose driver, so that I know the thing I'm invoking is NOT going to behave the way lldb normally behaves.

Jim

Personally I’m a big fan of this model. Sort of like what git does, where you have different install files in different scopes, and it starts by parsing the global config first, then the user-level config, then the local config, overwriting values in the process.

So in Ted’s scenario, for example, maybe lldb could load ~/.lldbinit first, then look in the same folder as the executable for a .lldbinit and load it second.

Actually after a quick glance at the code, maybe it already does something like this? I hadn’t really dug into the way .lldbinit works before.

First off, the fact that git does something at the user-interface level seems rather an argument against doing it that for...

But also, in git's case it makes some sense to have global configurations, for instance to implement policies for everybody on a team, or something like that. But I don't see that strong a use case for that for lldb. Plus, the particular instance seems another counter-argument, since as I understand it the Hexagon init is going to change the behavior of lldb significantly. I would rather if that happen I know this is not vanilla lldb, because it is called "lldb-hex" or something...

Jim

There isn't a global lldbinit. lldb handles the init files like gdb did, if you do:

$ lldb <SomeFile>

then we do:

command source ~/.lldbinit-<DRIVER_EXECUTABLE_NAME>

if that exists, or:

command source ~/.lldbinit

file <SomeFile>

command source <CWD>/.lldbinit

The latter is so you can do tricky things like have your build process stick a .lldbinit that is appropriate to debugging THAT binary somewhere, and then cd there before running the debug session.

Jim

All the llvm tools in our installation have "hexagon-" prepended too them, so lldb is really hexagon-lldb.

Right now we use a wrapper script that looks something like this:

DIR=$(dirname $0)
export PYTHONHOME=$DIR/..
exec $DIR/hexagon-lldb-3.5.0 -o "command script import $DIR/hexagon_utils.py" "$@"

I'd like to replace the '-o "command script import $DIR/hexagon_utils.py"' with something that gets loaded automatically, so I can set PYTHONHOME and just run hexagon-lldb-3.5.0 and get my utility scripts. I can do this with custom drivers (1 for lldb, 1 for lldb-mi), but I'd hoped to stay as close to the upstream code as possible.

If this is for the PYTHONHOME stuff, then could we solve this in a way similar to what I suggested on IRC? Basically, for Windows specifically, we either assume that python is installed into a specific location relative to the lldb executable, or look for an environment variable like LLDB_PYTHON_LOCATION (different from PYTHONHOME since that would be used by other installed versions of python as well).

Then you should be able to have a ".lldbinit-hexagon-lldb" file that will get sourced. The main questions is if there is a good place to put this file when there are no bundles on linux or windows. On LLDB, the LLDB.framework contains everything we need (headers, resources, support binaries (debugserver, lldb-gdbserver lldb-platform, darwin-debug, and more) and the lldb python modules.

Maybe we can start paving the way for bundles on windows and linux with LLDB?

Greg

I’m not too familiar with how bundles work on MacOSX, but I think having LLDB just be able to know where to look without any special configuration is the way to go. It sounds like that’s what bundles solve. In the interim I think a possible solution for this on Windows is to just have LLDB look for LLDB_PYTHON_HOME. If it finds it, it will expect to find Python there and call PySys_SetPath appropriately during initialization. If it doesn’t, then it’s up to the user to have a properly configured PYTHONHOME and PYTHONPATH.

This way, on hexagon machines, installation of LLDB could involve setting this environment variable to the appropriate value. It would also be useful for non-hexagon Windows as well, because the current state of affairs requires you to have a bunch of environment variables pointing to the correct locations on your disk, which is not very user friendly and a source of error when trying to build on Windows.

I’ve got a patch that’s being reviewed internally that lets you set cmake variables to set the default PYTHONHOME and PYTHONPATH. These get passed in on the compile line and are used to set PYTHONHOME and PYTHONPATH, if they’re not in the environment. I hope to have it done and up on phabricator soon. It works on Windows and Linux, and I assume other Unix variants.

that still doesn’t address your initial concern does it? Otherwise there would be no need for the global lldbinit file. What im suggesting removes the need for a PYTHONHOME and PYTHONPATH entirely, and instead relies on a one-time system setting.

Right. If the community doesn’t want an installation-global lldbinit file, I’ll just modify the drivers as suggested earlier.

To replace my wrapper scripts, I need that, and a way to find the python install automatically. I’ve just about got the 2nd one done J