diff --git a/compiler/rustc_trait_selection/src/traits/select/mod.rs b/compiler/rustc_trait_selection/src/traits/select/mod.rs
index 9e2d0657029fd..b4b367fe7c190 100644
--- a/compiler/rustc_trait_selection/src/traits/select/mod.rs
+++ b/compiler/rustc_trait_selection/src/traits/select/mod.rs
@@ -741,39 +741,7 @@ impl<'cx, 'tcx> SelectionContext<'cx, 'tcx> {
         if reached_depth >= stack.depth {
             debug!(?result, "CACHE MISS");
             self.insert_evaluation_cache(param_env, fresh_trait_pred, dep_node, result);
-
-            stack.cache().on_completion(
-                stack.dfn,
-                |fresh_trait_pred, provisional_result, provisional_dep_node| {
-                    // Create a new `DepNode` that has dependencies on:
-                    // * The `DepNode` for the original evaluation that resulted in a provisional cache
-                    // entry being crated
-                    // * The `DepNode` for the *current* evaluation, which resulted in us completing
-                    // provisional caches entries and inserting them into the evaluation cache
-                    //
-                    // This ensures that when a query reads this entry from the evaluation cache,
-                    // it will end up (transitively) depending on all of the incr-comp dependencies
-                    // created during the evaluation of this trait. For example, evaluating a trait
-                    // will usually require us to invoke `type_of(field_def_id)` to determine the
-                    // constituent types, and we want any queries reading from this evaluation
-                    // cache entry to end up with a transitive `type_of(field_def_id`)` dependency.
-                    //
-                    // By using `in_task`, we're also creating an edge from the *current* query
-                    // to the newly-created `combined_dep_node`. This is probably redundant,
-                    // but it's better to add too many dep graph edges than to add too few
-                    // dep graph edges.
-                    let ((), combined_dep_node) = self.in_task(|this| {
-                        this.tcx().dep_graph.read_index(provisional_dep_node);
-                        this.tcx().dep_graph.read_index(dep_node);
-                    });
-                    self.insert_evaluation_cache(
-                        param_env,
-                        fresh_trait_pred,
-                        combined_dep_node,
-                        provisional_result.max(result),
-                    );
-                },
-            );
+            stack.cache().on_completion(stack.dfn);
         } else {
             debug!(?result, "PROVISIONAL");
             debug!(
@@ -782,13 +750,7 @@ impl<'cx, 'tcx> SelectionContext<'cx, 'tcx> {
                 fresh_trait_pred, stack.depth, reached_depth,
             );
 
-            stack.cache().insert_provisional(
-                stack.dfn,
-                reached_depth,
-                fresh_trait_pred,
-                result,
-                dep_node,
-            );
+            stack.cache().insert_provisional(stack.dfn, reached_depth, fresh_trait_pred, result);
         }
 
         Ok(result)
@@ -2531,11 +2493,6 @@ struct ProvisionalEvaluation {
     from_dfn: usize,
     reached_depth: usize,
     result: EvaluationResult,
-    /// The `DepNodeIndex` created for the `evaluate_stack` call for this provisional
-    /// evaluation. When we create an entry in the evaluation cache using this provisional
-    /// cache entry (see `on_completion`), we use this `dep_node` to ensure that future reads from
-    /// the cache will have all of the necessary incr comp dependencies tracked.
-    dep_node: DepNodeIndex,
 }
 
 impl<'tcx> Default for ProvisionalEvaluationCache<'tcx> {
@@ -2578,7 +2535,6 @@ impl<'tcx> ProvisionalEvaluationCache<'tcx> {
         reached_depth: usize,
         fresh_trait_pred: ty::PolyTraitPredicate<'tcx>,
         result: EvaluationResult,
-        dep_node: DepNodeIndex,
     ) {
         debug!(?from_dfn, ?fresh_trait_pred, ?result, "insert_provisional");
 
@@ -2604,10 +2560,7 @@ impl<'tcx> ProvisionalEvaluationCache<'tcx> {
             }
         }
 
-        map.insert(
-            fresh_trait_pred,
-            ProvisionalEvaluation { from_dfn, reached_depth, result, dep_node },
-        );
+        map.insert(fresh_trait_pred, ProvisionalEvaluation { from_dfn, reached_depth, result });
     }
 
     /// Invoked when the node with dfn `dfn` does not get a successful
@@ -2633,8 +2586,7 @@ impl<'tcx> ProvisionalEvaluationCache<'tcx> {
     /// Invoked when the node at depth `depth` completed without
     /// depending on anything higher in the stack (if that completion
     /// was a failure, then `on_failure` should have been invoked
-    /// already). The callback `op` will be invoked for each
-    /// provisional entry that we can now confirm.
+    /// already).
     ///
     /// Note that we may still have provisional cache items remaining
     /// in the cache when this is done. For example, if there is a
@@ -2655,19 +2607,26 @@ impl<'tcx> ProvisionalEvaluationCache<'tcx> {
     /// would be 2, representing the C node, and hence we would
     /// remove the result for D, which has DFN 3, but not the results for
     /// A and B, which have DFNs 0 and 1 respectively).
-    fn on_completion(
-        &self,
-        dfn: usize,
-        mut op: impl FnMut(ty::PolyTraitPredicate<'tcx>, EvaluationResult, DepNodeIndex),
-    ) {
+    ///
+    /// Note that we *do not* attempt to cache these cycle participants
+    /// in the evaluation cache. Doing so would require carefully computing
+    /// the correct `DepNode` to store in the cache entry:
+    /// cycle participants may implicitly depend on query results
+    /// related to other participants in the cycle, due to our logic
+    /// which examines the evaluation stack.
+    ///
+    /// We used to try to perform this caching,
+    /// but it lead to multiple incremental compilation ICEs
+    /// (see #92987 and #96319), and was very hard to understand.
+    /// Fortunately, removing the caching didn't seem to
+    /// have a performance impact in practice.
+    fn on_completion(&self, dfn: usize) {
         debug!(?dfn, "on_completion");
 
         for (fresh_trait_pred, eval) in
             self.map.borrow_mut().drain_filter(|_k, eval| eval.from_dfn >= dfn)
         {
             debug!(?fresh_trait_pred, ?eval, "on_completion");
-
-            op(fresh_trait_pred, eval.result, eval.dep_node);
         }
     }
 }