[PATCH] D92155: Load plugins when creating a CompilerInvocation.

[Yafei Liu via Phabricator via cfe-commits]

psionic12 added a comment.

> This is exactly the problem, filenames specified in *clang* flags are *supposed* to be read from the VFS. (In practice this probably just means we'd need to disable this feature in environments where VFS is used)

I don't thinks so, plugins are loaded by using `-fplugin=xxxx.so`, this is a flag being parsed and saved in `FrontendOptions::Plugins`, which is a string vector. As you can see, `xxx.so` is treat as a string, not a file, vfs didn't load it to memory like normal files (for example `main.cpp`) . So I think whether VFS is used or not has nothing to do with plugin loading procedure.

> Sure. TL;DR is: clangd flags are configured by the user, user can be fully responsible for security/stability.
> clang flags are configured by the project. If they're bad, we can e.g. give bad diagnostics, but can't crash or compromise security.
>
> More detail:
>
> In the simplest possible case, clangd is configured as follows:
>
> 1. user downloads clangd binary
> 2. user installs an LSP plugin for their editor, and configures the plugin to use /usr/bin/clangd for C++ files. clangd starts when the editor does
> 3. the build system for $PROJECT generates $PROJECT/compile_commands.json
> 4. when the user opens $PROJECT/src/foo.cpp in the editor, it notifies clangd. clangd searches for $PROJECT/compile_commands.json, finds the clang arguments, and uses them to parse foo.cpp
>
> *clangd* command-line flags would be added explicitly by the user at step 2. We can reasonably ask the user to be aware/responsible for security/stability implications of doing this, including with their particular clangd version. We can also ask them to run `clangd --check` without the plugin flag to test whether the plugin is causing a stability problem.
>
> *clang* command-line flags are added implicitly in step 3. Or they could simply be checked into the repository - nothing ensures they were generated locally by the build system. The point is in typical usage they are not controlled by the user directly, and from a security perspective are not trusted (as safely opening files from untrusted repos is a reasonable expectation). So if we're loading plugins based on instructions in clang command-line flags, clangd bears most of the responsibility for making sure that's safe and correct (and I don't see a way to do that).

I got you point on this, using a clangd command-line flags means users are aware of the the plugin loading. Currently my idea is that if a plugin is listed in the compile_command.json, it will only be loaded if a flag with the same plugin name passed in from clangd command-line.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D92155/new/

https://reviews.llvm.org/D92155