Rollup merge of #108555 - Zoxc:par-fix, r=cjgillot

Fix a race in the query system

This fixes an issue where in between the `job` removal and `complete` call the query neither has a job nor a result, allowing another thread to start executing it again.

r? ``@cjgillot``

Author: matthiaskrgr

compiler/rustc_query_system/src/query/plumbing.rs

@@ -25,7 +25,6 @@ use std::collections::hash_map::Entry;
 use std::fmt::Debug;
 use std::hash::Hash;
 use std::mem;
-use std::ptr;
 use thin_vec::ThinVec;
 
 use super::QueryConfig;
@@ -250,13 +249,16 @@ where
     where
         C: QueryCache<Key = K>,
     {
-        // We can move out of `self` here because we `mem::forget` it below
-        let key = unsafe { ptr::read(&self.key) };
+        let key = self.key;
         let state = self.state;
 
         // Forget ourself so our destructor won't poison the query
         mem::forget(self);
 
+        // Mark as complete before we remove the job from the active state
+        // so no other thread can re-execute this query.
+        cache.complete(key, result, dep_node_index);
+
         let job = {
             #[cfg(parallel_compiler)]
             let mut lock = state.active.get_shard_by_value(&key).lock();
@@ -267,7 +269,6 @@ where
                 QueryResult::Poisoned => panic!(),
             }
         };
-        cache.complete(key, result, dep_node_index);
 
         job.signal_complete();
     }