On 7/29/2024 6:41 PM, Bruce Richardson wrote:
On Mon, Jul 29, 2024 at 06:16:48PM +0200, Burakov, Anatoly wrote:
On 7/29/2024 4:30 PM, Bruce Richardson wrote:
On Mon, Jul 29, 2024 at 02:05:52PM +0100, Anatoly Burakov wrote:
A lot of developers use Visual Studio Code as their primary IDE. This
script generates a configuration file for VSCode that sets up basic build
tasks, launch tasks, as well as C/C++ code analysis settings that will
take into account compile_commands.json that is automatically generated
by meson.
Files generated by script:
- .vscode/settings.json: stores variables needed by other files
- .vscode/tasks.json: defines build tasks
- .vscode/launch.json: defines launch tasks
- .vscode/c_cpp_properties.json: defines code analysis settings
The script uses a combination of globbing and meson file parsing to
discover available apps, examples, and drivers, and generates a
project-wide settings file, so that the user can later switch between
debug/release/etc. configurations while keeping their desired apps,
examples, and drivers, built by meson, and ensuring launch configurations
still work correctly whatever the configuration selected.
This script uses whiptail as TUI, which is expected to be universally
available as it is shipped by default on most major distributions.
However, the script is also designed to be scriptable and can be run
without user interaction, and have its configuration supplied from
command-line arguments.
Signed-off-by: Anatoly Burakov <anatoly.bura...@intel.com>
---
Just was trying this out, nice script, thanks.
Thanks for the feedback! Comments below.
Initial thoughts concerning the build directory:
- the script doesn't actually create the build directory, so there is no
guarantee that the build directory created will have the same parameters
as that specified in the script run. I'd suggest in the case where the
user runs the script and specifies build settings, that the build
directory is then configured using those settings.
I'm not sure I follow.
The script creates a command for VSCode to create a build directory using
configuration the user has supplied at script's run time. The directory is
not created by the script, that is the job of meson build system. This
script is merely codifying commands for meson to do that, with the
expectation that the user is familiar with VSCode workflow and will go
straight to build commands anyway, and will pick one of them. Are you
suggesting running `meson setup` right after?
Yes, that is what I was thinking, based on the assumption that running
"meson setup" should be a once-only task. I suppose overall it's a
different workflow to what you have - where you run meson setup repeatedly
each time you change a build type. My thinking for the approach to take
here is that your script firstly asks for a build directory, and then:
* if build-dir exists, pull settings you need from there, such as build
type and the apps being built.
* if not existing, ask for config settings as you do now, and then run
meson setup to create the build dir.
I guess the disconnect comes from the fact that I treat meson build
directories as transient and disposable and never hesitate to wipe them
and start over, whereas you seem to favor persistence. Since 99% of the
time I'm using heavily reduced builds anyway (e.g. one app, one driver),
repeatedly doing meson setup doesn't really hinder me in any way, but I
suppose it would if I had more meaty builds.
I think we also have a slightly different view of what this script
should be - I envisioned a "one-stop-shop" for "open freshly cloned DPDK
directory, run one script and off you go" (which stems from the fact
that I heavily rely on Git Worktrees [1] in my workflow, so having an
untouched source directory without any configuration in it is something
I am constantly faced with), while you seem to favor picking up existing
meson build and building a VSCode configuration around it.
I can see point in this, and I guess I actually unwittingly rolled two
scripts into one - a TUI meson frontend, and a VSCode configuration
generator. Perhaps it should very well be two scripts, not one? Because
honestly, having something like what I built for TUI (the meson
configuration frontend) is something I missed a lot and something I
always wished our long-gone `setup.sh` script had, but didn't, and now
with meson it's much simpler to do but we still don't have anything like
that. Maybe this is our opportunity to provide a quicker "quick start"
script, one that I could run and configure meson build without having to
type anything. WDYT?
Thereafter, the source for all build settings is not in vscode, but in
meson, and you just use "meson configure" from vscode to tweak whatever
needs tweaking, without affecting any other settings. Since you don't
affect any other settings there is no need to record what they should be.
Why would I be using meson configure from vscode then? I mean, the whole
notion of having tasks in VSCode is so that I define a few common
configurations and never touch them until I'm working on something else.
If I have to type in configure commands anyway (because I would have to,
to adjust configuration in meson), I might as well do so using terminal,
and avoid dealing with meson from VSCode entirely?
(technically, there's a way to read arguments from a popup window - I
suppose we could have this step recorded as a VSCode task)
Assuming we do that, it would actually then be possible to adjust launch
tasks to only include *actual* built apps/examples (as well as infer things
like platform, compiler etc.), as that's one weakness of my current "flying
blind" approach, so I wouldn't be opposed to adding an extra step here, just
want to make sure I understand what you're saying correctly.
- On the other hand, when the build directory already exists - I think the
script should pull all settings from there, rather than prompting the
user.
That can be done, however, my own workflow has been that I do not ever keep
build directories inside my source directory, so it would not be possible to
pick up directories anywhere but the source directory.
Why is that, and how does it work now, for e.g. getting the
compile_commands.json file from your build folder?
Right now, at configuration time I store build directory in
settings.json, and all of the other tasks (build, launch, C++ analysis)
pick it up via ${config:...} variable. This is why I suggested having a
notion of "active configuration" - if we just rewrite that variable,
both ninja and launch tasks can be switched to a different build
directory without actually having to create new tasks. More on that
below, as that raises a few questions.
As for why, it's a personal preference - it's annoying to have the
build/ directory in my file view (it's possible to hide it, I guess I
could automatically add a setting to do that in settings.json), and it
interferes with e.g. source-tree-wide searches and stuff like that. This
isn't a hard requirement though, I can switch to in-tree builds and
automatic directory hiding if it means more automation :)
I also think from the point of view of the script it would be easier to
start from known quantities rather than guess what user was trying to do
from current configuration, but I guess a few common-sense heuristics should
suffice for most use cases, such as e.g. inferring debug builds.
What you need depends on whether you want to keep running "meson setup" -
which means you need to track all settings - or want to use "meson
configure" where you don't really need to track much at all.
I guess I was aiming for the former, but doing only the latter makes
sense if we assume we want to separate setup from VSCode config
generation (which it seems like we're heading that way).
- I'm not sure I like the idea for reconfiguring of just removing the build
directory and doing a whole meson setup command all over again. This
seems excessive and also removes the possibility of the user having made
changes in config to the build dir without re-running the whole config
script. For example, having tweaked the LTO setting, or the
instruction_set_isa_setting. Rather than deleting it and running meson
setup, it would be better to use "meson configure" to adjust the one
required setting and let ninja figure out how to propagate that change.
That saves the script from having to track all meson parameters itself.
Last I checked, meson doesn't have a command that would "setup or configure
existing" a directory, it's either "set up new one" or "configure existing
one". I guess we could set up a fallback of "configure || setup".
This goes back to the whole "create build directory after running the
script" option. If the script creates the build dir, the vscode commands
never need to use meson setup, only ever meson configure.
Sure, and like I suggested above, it doesn't even have to be *this*
script, it can be another, a Python-based TUI reimplementation of
setup.sh :)
- Finally, and semi-related, this script assumes that the user does
everything in a single build directory. Just something to consider, but
my own workflow till now has tended to keep multiple build directories
around, generally a "build" directory, which is either release or
debugoptimized type, and a separate "build-debug" directory + occasionally
others for build testing. When doing incremental builds, the time taken to
do two builds following a change is a lot less noticable than the time taken
for periodic switches of a single build directory between debug and release
mode.
The problem with that approach is the launch tasks, because a launch task
can only ever point to one executable, so if you have multiple build
directories, you'll have to have multiple launch tasks per app/example. I
guess we can either tag them (e.g. "Launch dpdk-testpmd [debug]", "Launch
dpdk-testpmd [asan]" etc.), or use some kind of indirection to "select
active build configuration" (e.g. have one launch task but overwrite
${config:BUILDDIR} after request for configuration, so that launch tasks
would pick up actual executable path at run time from settings). I would
prefer the latter to be honest, as it's much easier to drop a script into
./vscode and run it together with "configure" command to switch between
different build/launch configurations. What do you think?
I think I'd prefer the former actually - to have explicit tasks always
listed for debug and release builds.
Not a big deal for me either way, I'll just hack in the extra tasks as I
need them, so it's a low-priority support item for me.
There's another issue though: code analysis. If you have multiple build
directories, your C++ analysis settings (the
.vscode/c_cpp_properties.json file) can only ever use one specific
compile_commands.json from a specific build directory. I think if we
were to support having multiple build dirs, we would *have to* implement
something like "switch active configuration" thing anyway.
Whether we add tagged launch tasks is honestly a problem to be solved,
because on the one hand we want them to be generated automatically
(ideally after configure step), and on the other we also want to not
interfere with any custom configuration the user has already added (such
as e.g. command line arguments to existing launch tasks). We'll probably
have to do config file parsing and updating the configuration, rather
than regenerating it each time. Python's JSON also doesn't support
comments, so for any comments that were added to configurations, we'd
either lose them, or find a way to reinsert them post-generation.
Final thoughts on usability:
- Please don't write gdbsudo to /usr/local/bin without asking the user
first. Instead I think it should default to $HOME/.local/bin, but with a
prompt for the user to specify a path.
It's not creating anything, it's just printing out a snippet, which, if run
by user, would do that - the implication is obviously that the user may
correct it if necessary. The script actually picks up path to `gdbsudo` from
`which` command, so if the user puts their command to $HOME/.local/bin or
something, it would get picked up if it's in PATH. I see your point about
maybe suggesting using a home directory path instead of a system wide path,
I can change that.
Yep, thanks, and thanks for the explanation.
BTW: even if the user is running as non-root user, they don't always need
to use sudo (I set up my system to not need it for running DPDK). I see
there is a cmdline option for "no-gdbsudo" but I think you should make that
accessible via TUI also if non-root.
Sure, all of that can be done.
And for the cmdline parameters, how about shortening them to "--sudo" and
"--nosudo". For debug vs release builds you may want to have the latter run
without gdb at all, just with or without sudo.]
I honestly never run anything outside gdb, but that's a consequence of
me not really working on things that routinely require it (e.g.
performace code). We can add that no problem though.
- While I realise your primary concern here is an interactive script, I'd
tend towards requiring a cmdline arg to run in interactive mode and
instead printing the help usage when run without parameters. Just a
personal preference on my part though.
I found it to be much faster to pick my targets, apps etc. using a couple of
interactive windows than to type out parameters I probably don't even
remember ahead of time (especially build configurations!), and I believe
it's more newbie-friendly that way, as I imagine very few people will want
to learn arguments for yet-another-script just to start using VSCode. It
would be my personal preference to leave it as default-to-TUI, but maybe
recognizing a widely used `-i` parameter would be a good compromise for
instant familiarity.
Ok. There's always a -h option for me to get the cmdline parameters.
I also think if the script is ok with working off an existing build
directory (or directories!), and only prompting for that, it would remove
for me the real necessity of asking for a more cmdline-fieldly version.
It sounds like this would really be something that a setup script would
do better than a VSCode config generator.
So, assuming we want to move setup steps to another script and
concentrate on VSCode configuration exclusively, my thinking of how it
would work is as follows:
1) Assume we want multiple build directories, suggest automatically
picking them up from source directory but support specifying one or more
from command line arguments (or TUI, although I suspect if setup moves
to a separate script, there's very little need for TUI in VSCode config
generator - it should be a mostly mechanical process at that point)
2) Now that we track multiple build directories, we can store them in a
YAML file or something (e.g. .vscode/.dpdk_builds.yaml), and create
tasks to switch between them as "active" to support e.g. different code
analysis settings and stuff like that
3) All build tasks can work off "active configuration" which means we
don't need multiple compile tasks, but for launch tasks we may need
different ones because different build dirs may have different launch tasks
Let's assume user just ran `meson configure` and changed something about
one of their configurations. What do you think should happen next? I
mean, if they added/removed an app/example to the build, we can detect
that and auto-generate (or remove) a launch task, for example? Or do you
think it should be a manual step, e.g. user should explicitly request
regenerating/updating launch tasks? Maybe there should be an --update
flag, to indicate that we're not creating new configurations but merely
refreshing existing ones?
[1] https://git-scm.com/docs/git-worktree
--
Thanks,
Anatoly
