Compute pipelines never freed at runtime, leaking memory

[jimblandy]

See also #2546. The following program drops everything but the instance, and yet calling instance.poll_all(true) never causes the compute pipeline to be freed at the hal level:

use anyhow::Result; // anyhow = "1"
use futures_executor::block_on; // futures-executor = "0.3"

fn main() -> Result<()> {
    env_logger::init();
    
    let instance = wgpu::Instance::new(wgpu::Backends::VULKAN);

    {
        let (device, queue) = block_on(async {
            let adapter = instance.request_adapter(&wgpu::RequestAdapterOptions::default())
                .await
                .expect("request_adapter returned None");
            adapter.request_device(&wgpu::DeviceDescriptor {
                label: Some("play-wgpu device"),
                features: wgpu::Features::default(),
                limits: wgpu::Limits::default(),
            },
                                   None).await
        })?;

        let module = device.create_shader_module(&wgpu::ShaderModuleDescriptor {
            label: Some("play-wgpu shader module"),
            source: wgpu::ShaderSource::Wgsl("@stage(compute) @workgroup_size(32) fn main() {}".into())
        });

        let pipeline_layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
            label: Some("play-wgpu pipeline layout descriptor"),
            bind_group_layouts: &[],
            push_constant_ranges: &[],
        });
        
        let compute_pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
            label: Some("play-wgpu compute pipeline"),
            layout: Some(&pipeline_layout),
            module: &module,
            entry_point: "main",
        });

        drop((compute_pipeline, queue));
    }
    
    instance.poll_all(true);
    log::info!("called instance.poll_all with everything out of scope");
    // compute pipeline has not yet been freed at this point

    Ok(())
}

[jimblandy]

Global::compute_pipeline_drop takes the LifeGuard's RefCount and drops it, and adds the pipeline's id to the device's suspected_resources, but because the pipeline was never added to the device's TrackerSet, when triage_suspected calls trackers.compute_pipes.remove_abandoned(id), it always gets false, so it never ends up adding the raw pipeline to free_resources (nor anyone's last_resources).

[jimblandy]

Similarly for render pipelines.