diff --git a/Compiler/src/typeinfer.jl b/Compiler/src/typeinfer.jl
index 1afe4c297e5ba..43b743982ea89 100644
--- a/Compiler/src/typeinfer.jl
+++ b/Compiler/src/typeinfer.jl
@@ -399,7 +399,7 @@ function transform_result_for_cache(interp::AbstractInterpreter, result::Inferen
     if isa(src, OptimizationState)
         opt = src
         inlining_cost = compute_inlining_cost(interp, result, opt.optresult)
-        discard_optimized_result(interp, opt, inlining_cost) && return nothing
+        discard_optimized_result(interp, opt, inlining_cost, result.ipo_effects) && return nothing
         src = ir_to_codeinf!(opt)
     end
     if isa(src, CodeInfo)
@@ -409,7 +409,7 @@ function transform_result_for_cache(interp::AbstractInterpreter, result::Inferen
     return src
 end
 
-function discard_optimized_result(interp::AbstractInterpreter, opt#=::OptimizationState=#, inlining_cost#=::InlineCostType=#)
+function discard_optimized_result(interp::AbstractInterpreter, opt#=::OptimizationState=#, inlining_cost#=::InlineCostType=#, effects::Effects)
     may_discard_trees(interp) || return false
     return inlining_cost == MAX_INLINE_COST
 end
diff --git a/src/aotcompile.cpp b/src/aotcompile.cpp
index d687f44808409..5cdecda9d8582 100644
--- a/src/aotcompile.cpp
+++ b/src/aotcompile.cpp
@@ -675,20 +675,6 @@ void *jl_create_native_impl(jl_array_t *methods, LLVMOrcThreadSafeModuleRef llvm
     fargs[0] = (jl_value_t*)codeinfos;
     void *data = jl_emit_native(codeinfos, llvmmod, &cgparams, external_linkage);
 
-    // examine everything just emitted and save it to the caches
-    if (!external_linkage) {
-        for (size_t i = 0, l = jl_array_nrows(codeinfos); i < l; i++) {
-            jl_value_t *item = jl_array_ptr_ref(codeinfos, i);
-            if (jl_is_code_instance(item)) {
-                // now add it to our compilation results
-                jl_code_instance_t *codeinst = (jl_code_instance_t*)item;
-                jl_code_info_t *src = (jl_code_info_t*)jl_array_ptr_ref(codeinfos, ++i);
-                assert(jl_is_code_info(src));
-                jl_add_codeinst_to_cache(codeinst, src);
-            }
-        }
-    }
-
     // move everything inside, now that we've merged everything
     // (before adding the exported headers)
     ((jl_native_code_desc_t*)data)->M.withModuleDo([&](Module &M) {
diff --git a/src/gf.c b/src/gf.c
index 6583262798806..1a51f7cd817ec 100644
--- a/src/gf.c
+++ b/src/gf.c
@@ -2836,30 +2836,10 @@ void jl_read_codeinst_invoke(jl_code_instance_t *ci, uint8_t *specsigflags, jl_c
 
 jl_method_instance_t *jl_normalize_to_compilable_mi(jl_method_instance_t *mi JL_PROPAGATES_ROOT);
 
-JL_DLLEXPORT void jl_add_codeinst_to_cache(jl_code_instance_t *codeinst, jl_code_info_t *src)
-{
-    assert(jl_is_code_info(src));
-    jl_method_instance_t *mi = jl_get_ci_mi(codeinst);
-    if (jl_generating_output() && jl_is_method(mi->def.method) && jl_atomic_load_relaxed(&codeinst->inferred) == jl_nothing) {
-        jl_value_t *compressed = jl_compress_ir(mi->def.method, src);
-        // These should already be compatible (and should be an assert), but make sure of it anyways
-        if (jl_is_svec(src->edges)) {
-            jl_atomic_store_release(&codeinst->edges, (jl_svec_t*)src->edges);
-            jl_gc_wb(codeinst, src->edges);
-        }
-        jl_atomic_store_release(&codeinst->debuginfo, src->debuginfo);
-        jl_gc_wb(codeinst, src->debuginfo);
-        jl_atomic_store_release(&codeinst->inferred, compressed);
-        jl_gc_wb(codeinst, compressed);
-    }
-}
-
-
 JL_DLLEXPORT void jl_add_codeinst_to_jit(jl_code_instance_t *codeinst, jl_code_info_t *src)
 {
     assert(jl_is_code_info(src));
     jl_emit_codeinst_to_jit(codeinst, src);
-    jl_add_codeinst_to_cache(codeinst, src);
 }
 
 jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t world)
diff --git a/src/julia_internal.h b/src/julia_internal.h
index 9966d36027473..fd8720b7f1701 100644
--- a/src/julia_internal.h
+++ b/src/julia_internal.h
@@ -685,7 +685,6 @@ JL_DLLEXPORT jl_method_instance_t *jl_get_unspecialized(jl_method_t *def JL_PROP
 JL_DLLEXPORT void jl_read_codeinst_invoke(jl_code_instance_t *ci, uint8_t *specsigflags, jl_callptr_t *invoke, void **specptr, int waitcompile);
 JL_DLLEXPORT jl_method_instance_t *jl_method_match_to_mi(jl_method_match_t *match, size_t world, size_t min_valid, size_t max_valid, int mt_cache);
 JL_DLLEXPORT void jl_add_codeinst_to_jit(jl_code_instance_t *codeinst, jl_code_info_t *src);
-JL_DLLEXPORT void jl_add_codeinst_to_cache(jl_code_instance_t *codeinst, jl_code_info_t *src);
 
 JL_DLLEXPORT jl_code_instance_t *jl_new_codeinst_uninit(jl_method_instance_t *mi, jl_value_t *owner);
 JL_DLLEXPORT jl_code_instance_t *jl_new_codeinst(
diff --git a/src/staticdata.c b/src/staticdata.c
index c96faf5a71a54..4c1489770aa7e 100644
--- a/src/staticdata.c
+++ b/src/staticdata.c
@@ -863,40 +863,60 @@ static void jl_insert_into_serialization_queue(jl_serializer_state *s, jl_value_
         }
         goto done_fields; // for now
     }
-    if (s->incremental && jl_is_method_instance(v)) {
+    if (jl_is_method_instance(v)) {
         jl_method_instance_t *mi = (jl_method_instance_t*)v;
-        jl_value_t *def = mi->def.value;
-        if (needs_uniquing(v, s->query_cache)) {
-            // we only need 3 specific fields of this (the rest are not used)
-            jl_queue_for_serialization(s, mi->def.value);
-            jl_queue_for_serialization(s, mi->specTypes);
-            jl_queue_for_serialization(s, (jl_value_t*)mi->sparam_vals);
-            goto done_fields;
-        }
-        else if (jl_is_method(def) && jl_object_in_image(def)) {
-            // we only need 3 specific fields of this (the rest are restored afterward, if valid)
-            // in particular, cache is repopulated by jl_mi_cache_insert for all foreign function,
-            // so must not be present here
-            record_field_change((jl_value_t**)&mi->backedges, NULL);
-            record_field_change((jl_value_t**)&mi->cache, NULL);
+        if (s->incremental) {
+            jl_value_t *def = mi->def.value;
+            if (needs_uniquing(v, s->query_cache)) {
+                // we only need 3 specific fields of this (the rest are not used)
+                jl_queue_for_serialization(s, mi->def.value);
+                jl_queue_for_serialization(s, mi->specTypes);
+                jl_queue_for_serialization(s, (jl_value_t*)mi->sparam_vals);
+                goto done_fields;
+            }
+            else if (jl_is_method(def) && jl_object_in_image(def)) {
+                // we only need 3 specific fields of this (the rest are restored afterward, if valid)
+                // in particular, cache is repopulated by jl_mi_cache_insert for all foreign function,
+                // so must not be present here
+                record_field_change((jl_value_t**)&mi->backedges, NULL);
+                record_field_change((jl_value_t**)&mi->cache, NULL);
+            }
+            else {
+                assert(!needs_recaching(v, s->query_cache));
+            }
+            // n.b. opaque closures cannot be inspected and relied upon like a
+            // normal method since they can get improperly introduced by generated
+            // functions, so if they appeared at all, we will probably serialize
+            // them wrong and segfault. The jl_code_for_staged function should
+            // prevent this from happening, so we do not need to detect that user
+            // error now.
         }
-        else {
-            assert(!needs_recaching(v, s->query_cache));
+        // don't recurse into all backedges memory (yet)
+        jl_value_t *backedges = get_replaceable_field((jl_value_t**)&mi->backedges, 1);
+        if (backedges) {
+            jl_queue_for_serialization_(s, (jl_value_t*)((jl_array_t*)backedges)->ref.mem, 0, 1);
+            size_t i = 0, n = jl_array_nrows(backedges);
+            while (i < n) {
+                jl_value_t *invokeTypes;
+                jl_code_instance_t *caller;
+                i = get_next_edge((jl_array_t*)backedges, i, &invokeTypes, &caller);
+                if (invokeTypes)
+                    jl_queue_for_serialization(s, invokeTypes);
+            }
         }
-        // n.b. opaque closures cannot be inspected and relied upon like a
-        // normal method since they can get improperly introduced by generated
-        // functions, so if they appeared at all, we will probably serialize
-        // them wrong and segfault. The jl_code_for_staged function should
-        // prevent this from happening, so we do not need to detect that user
-        // error now.
-    }
-    if (s->incremental && jl_is_binding(v)) {
-        if (needs_uniquing(v, s->query_cache)) {
-            jl_binding_t *b = (jl_binding_t*)v;
+    }
+    if (jl_is_binding(v)) {
+        jl_binding_t *b = (jl_binding_t*)v;
+        if (s->incremental && needs_uniquing(v, s->query_cache)) {
             jl_queue_for_serialization(s, b->globalref->mod);
             jl_queue_for_serialization(s, b->globalref->name);
             goto done_fields;
         }
+        // don't recurse into backedges memory (yet)
+        jl_value_t *backedges = get_replaceable_field((jl_value_t**)&b->backedges, 1);
+        if (backedges) {
+            jl_queue_for_serialization_(s, (jl_value_t*)((jl_array_t*)backedges)->ref.mem, 0, 1);
+        }
     }
     if (s->incremental && jl_is_globalref(v)) {
         jl_globalref_t *gr = (jl_globalref_t*)v;
@@ -914,18 +934,20 @@ static void jl_insert_into_serialization_queue(jl_serializer_state *s, jl_value_
             assert(!jl_object_in_image((jl_value_t*)tn->wrapper));
         }
     }
-    if (s->incremental && jl_is_code_instance(v)) {
+    if (jl_is_code_instance(v)) {
         jl_code_instance_t *ci = (jl_code_instance_t*)v;
         jl_method_instance_t *mi = jl_get_ci_mi(ci);
-        // make sure we don't serialize other reachable cache entries of foreign methods
-        // Should this now be:
-        // if (ci !in ci->defs->cache)
-        //     record_field_change((jl_value_t**)&ci->next, NULL);
-        // Why are we checking that the method/module this originates from is in_image?
-        // and then disconnect this CI?
-        if (jl_object_in_image((jl_value_t*)mi->def.value)) {
-            // TODO: if (ci in ci->defs->cache)
-            record_field_change((jl_value_t**)&ci->next, NULL);
+        if (s->incremental) {
+            // make sure we don't serialize other reachable cache entries of foreign methods
+            // Should this now be:
+            // if (ci !in ci->defs->cache)
+            //     record_field_change((jl_value_t**)&ci->next, NULL);
+            // Why are we checking that the method/module this originates from is in_image?
+            // and then disconnect this CI?
+            if (jl_object_in_image((jl_value_t*)mi->def.value)) {
+                // TODO: if (ci in ci->defs->cache)
+                record_field_change((jl_value_t**)&ci->next, NULL);
+            }
         }
         jl_value_t *inferred = jl_atomic_load_relaxed(&ci->inferred);
         if (inferred && inferred != jl_nothing) { // disregard if there is nothing here to delete (e.g. builtins, unspecialized)
@@ -953,7 +975,7 @@ static void jl_insert_into_serialization_queue(jl_serializer_state *s, jl_value_
                 if (inferred == jl_nothing) {
                     record_field_change((jl_value_t**)&ci->inferred, jl_nothing);
                 }
-                else if (jl_is_string(inferred)) {
+                else if (s->incremental && jl_is_string(inferred)) {
                     // New roots for external methods
                     if (jl_object_in_image((jl_value_t*)def)) {
                         void **pfound = ptrhash_bp(&s->method_roots_index, def);
@@ -2572,6 +2594,35 @@ static void jl_prune_type_cache_linear(jl_svec_t *cache)
         jl_svecset(cache, ins++, jl_nothing);
 }
 
+static void jl_prune_mi_backedges(jl_array_t *backedges)
+{
+    if (backedges == NULL)
+        return;
+    size_t i = 0, ins = 0, n = jl_array_nrows(backedges);
+    while (i < n) {
+        jl_value_t *invokeTypes;
+        jl_code_instance_t *caller;
+        i = get_next_edge(backedges, i, &invokeTypes, &caller);
+        if (ptrhash_get(&serialization_order, caller) != HT_NOTFOUND)
+            ins = set_next_edge(backedges, ins, invokeTypes, caller);
+    }
+    jl_array_del_end(backedges, n - ins);
+}
+
+static void jl_prune_binding_backedges(jl_array_t *backedges)
+{
+    if (backedges == NULL)
+        return;
+    size_t i = 0, ins = 0, n = jl_array_nrows(backedges);
+    for (i = 0; i < n; i++) {
+        jl_value_t *b = jl_array_ptr_ref(backedges, i);
+        if (ptrhash_get(&serialization_order, b) != HT_NOTFOUND)
+            jl_array_ptr_set(backedges, ins, b);
+    }
+    jl_array_del_end(backedges, n - ins);
+}
+
+
 uint_t bindingkey_hash(size_t idx, jl_value_t *data);
 
 static void jl_prune_module_bindings(jl_module_t * m) JL_GC_DISABLED
@@ -3145,12 +3196,11 @@ static void jl_save_system_image_to_stream(ios_t *f, jl_array_t *mod_array,
             jl_queue_for_serialization(&s, global_roots_keyset);
             jl_serialize_reachable(&s);
         }
-        // step 1.5: prune (garbage collect) some special weak references from
-        // built-in type caches too
+        // step 1.5: prune (garbage collect) some special weak references known caches
         for (i = 0; i < serialization_queue.len; i++) {
             jl_value_t *v = (jl_value_t*)serialization_queue.items[i];
             if (jl_options.trim) {
-                if (jl_is_method(v)){
+                if (jl_is_method(v)) {
                     jl_method_t *m = (jl_method_t*)v;
                     jl_value_t *specializations_ = jl_atomic_load_relaxed(&m->specializations);
                     if (!jl_is_svec(specializations_))
@@ -3178,6 +3228,16 @@ static void jl_save_system_image_to_stream(ios_t *f, jl_array_t *mod_array,
                 jl_gc_wb(tn, jl_atomic_load_relaxed(&tn->cache));
                 jl_prune_type_cache_linear(jl_atomic_load_relaxed(&tn->linearcache));
             }
+            else if (jl_is_method_instance(v)) {
+                jl_method_instance_t *mi = (jl_method_instance_t*)v;
+                jl_value_t *backedges = get_replaceable_field((jl_value_t**)&mi->backedges, 1);
+                jl_prune_mi_backedges((jl_array_t*)backedges);
+            }
+            else if (jl_is_binding(v)) {
+                jl_binding_t *b = (jl_binding_t*)v;
+                jl_value_t *backedges = get_replaceable_field((jl_value_t**)&b->backedges, 1);
+                jl_prune_binding_backedges((jl_array_t*)backedges);
+            }
         }
     }