[Macros] Recovery after executable plugin crash

[rintaro]

When executable plugins crashed or somehow decided to exit, the compiler
should relaunch the plugin executable before sending another message.

• LoadedExecutablePlugin is now a process manager of the plugin executable.
• LoadedExecutabePlugin::PluginProcess represents the actual plugin process
• LoadedExecutablePlugin::spawnIfNeeded()

[rintaro]

@swift-ci Please smoke test

[rintaro]

@swift-ci Please smoke test

[rintaro]

@swift-ci Please smoke test

[rintaro]

@swift-ci Please smoke test

[rintaro]

These were unused declarations accidentally added in 0e31393

[rintaro]

s/data()/c_str()/g. Ditto for the callback.

[bnbarham]

error if ... isn't finished

[bnbarham]

You could have this as a local instead and only set it after it's valid (ie. after the spawnIfNeeded). Not crucial, but nice to have it not set unless it's valid IMO.

[bnbarham]

Someone could theoretically kill the process between where we spawn and sendMessage. It's handled below (ie. write would fail), so I'd probably just remove this.

[bnbarham]

We should probably change this message + write out a reproducer for the failure on read. On write... I'm not really sure what to add 😅. Not for this PR though

[rintaro]

This is not safe when the PluginRegistry is owned by ASTContext. Currently when ASTContext is to own a PluginRegistry, it adds a cleanup function like this->addCleanup([registry]{ delete registry; }), and cleanup functions are called in the order they are added. That means that the plugin registry and its plugins would be deleted before this callback is called. To resolve this:

• Instead of using cleanup callback for the plugin registry, add a fieldstd::unique_ptr<PluginRegistry> OwnedPluginRegistry to ASTContext::Implementation so it'd be alive until these callbacks are called
• or, call ASTContext cleanup callbacks in reverse order, like a stack

@bnbarham wdyt?

[rintaro]

Went with the former because I think we should reduce the use of addCleanup whenever possible.

[bnbarham]

Sounds good to me. I'm all for more managed pointers :). IMO the latter also makes sense as a general change.

[rintaro]

@swift-ci Please smoke test