Asyncify liveness analysis (#2890)

This finds out which locals are live at call sites that might pause/resume,
which is the set of locals we need to actually save/load. That is, if a local
is not alive at any call site in the function, then it's value doesn't need to
stay alive while sleeping.

This saves about 10% of locals that are saved/loaded, and about 1.5%
in final code size.

11 files changed, 1453 insertions, 854 deletions

diff --git a/scripts/fuzz_opt.py b/scripts/fuzz_opt.py
index 76d97bd2e..9c8af1106 100755
--- a/scripts/fuzz_opt.py
+++ b/scripts/fuzz_opt.py
@@ -545,10 +545,6 @@ class Asyncify(TestCaseHandler):
             print('ignoring due to pre-asyncify difference')
             return
 
-        # TODO: also something that actually does async sleeps in the code, say
-        # on the logging commands?
-        # --remove-unused-module-elements removes the asyncify intrinsics, which are not valid to call
-
         def do_asyncify(wasm):
             cmd = [in_bin('wasm-opt'), wasm, '--asyncify', '-o', 'async.t.wasm']
             # if we allow NaNs, running binaryen optimizations and then
diff --git a/src/passes/Asyncify.cpp b/src/passes/Asyncify.cpp
index 2708ae4e0..d4a788adb 100644
--- a/src/passes/Asyncify.cpp
+++ b/src/passes/Asyncify.cpp
@@ -291,6 +291,7 @@
 //      of their original range.
 //
 
+#include "cfg/liveness-traversal.h"
 #include "ir/effects.h"
 #include "ir/find_all.h"
 #include "ir/literal-utils.h"
@@ -1149,8 +1150,9 @@ struct AsyncifyLocals : public WalkerPass<PostWalker<AsyncifyLocals>> {
     if (!analyzer->needsInstrumentation(func)) {
       return;
     }
-    // Note the locals we want to preserve, before we add any more helpers.
-    numPreservableLocals = func->getNumLocals();
+    // Find the locals that we actually need to load and save: any local that is
+    // alive at a relevant call site must be handled, but others can be ignored.
+    findRelevantLiveLocals(func);
     // The new function body has a prelude to load locals if rewinding,
     // then the actual main body, which does all its unwindings by breaking
     // to the unwind block, which then handles pushing the call index, as
@@ -1198,16 +1200,61 @@ private:
   std::unique_ptr<AsyncifyBuilder> builder;
 
   Index rewindIndex;
-  Index numPreservableLocals;
   std::map<Type, Index> fakeCallLocals;
+  std::set<Index> relevantLiveLocals;
+
+  void findRelevantLiveLocals(Function* func) {
+    struct RelevantLiveLocalsWalker
+      : public LivenessWalker<RelevantLiveLocalsWalker,
+                              Visitor<RelevantLiveLocalsWalker>> {
+      // Basic blocks that have a possible unwind/rewind in them.
+      std::set<BasicBlock*> relevantBasicBlocks;
+
+      void visitCall(Call* curr) {
+        if (!currBasicBlock) {
+          return;
+        }
+        // Note blocks where we might unwind/rewind, all of which have a
+        // possible call to ASYNCIFY_CHECK_CALL_INDEX emitted right before the
+        // actual call.
+        // Note that each relevant original call was turned into a sequence of
+        // instructions, one of which is an if and then a call to this special
+        // intrinsic. We rely on the fact that if a local was live at the
+        // original call, it also would be in all that sequence of instructions,
+        // and in particular at the call we look for here (which is right before
+        // the call, and so anything that has its final use at the call is still
+        // live here).
+        if (curr->target == ASYNCIFY_CHECK_CALL_INDEX) {
+          relevantBasicBlocks.insert(currBasicBlock);
+        }
+      }
+    };
+
+    RelevantLiveLocalsWalker walker;
+    walker.walkFunctionInModule(func, getModule());
+    // The relevant live locals are ones that are alive at an unwind/rewind
+    // location. TODO look more precisely inside basic blocks, as one might stop
+    // being live in the middle
+    for (auto* block : walker.liveBlocks) {
+      if (walker.relevantBasicBlocks.count(block)) {
+        for (auto local : block->contents.start) {
+          relevantLiveLocals.insert(local);
+        }
+      }
+    }
+  }
 
   Expression* makeLocalLoading() {
-    if (numPreservableLocals == 0) {
+    if (relevantLiveLocals.empty()) {
       return builder->makeNop();
     }
     auto* func = getFunction();
+    auto numLocals = func->getNumLocals();
     Index total = 0;
-    for (Index i = 0; i < numPreservableLocals; i++) {
+    for (Index i = 0; i < numLocals; i++) {
+      if (!relevantLiveLocals.count(i)) {
+        continue;
+      }
       total += func->getLocalType(i).getByteSize();
     }
     auto* block = builder->makeBlock();
@@ -1216,7 +1263,10 @@ private:
     block->list.push_back(
       builder->makeLocalSet(tempIndex, builder->makeGetStackPos()));
     Index offset = 0;
-    for (Index i = 0; i < numPreservableLocals; i++) {
+    for (Index i = 0; i < numLocals; i++) {
+      if (!relevantLiveLocals.count(i)) {
+        continue;
+      }
       const auto& types = func->getLocalType(i).expand();
       SmallVector<Expression*, 1> loads;
       for (Index j = 0; j < types.size(); j++) {
@@ -1248,16 +1298,20 @@ private:
   }
 
   Expression* makeLocalSaving() {
-    if (numPreservableLocals == 0) {
+    if (relevantLiveLocals.empty()) {
       return builder->makeNop();
     }
     auto* func = getFunction();
+    auto numLocals = func->getNumLocals();
     auto* block = builder->makeBlock();
     auto tempIndex = builder->addVar(func, Type::i32);
     block->list.push_back(
       builder->makeLocalSet(tempIndex, builder->makeGetStackPos()));
     Index offset = 0;
-    for (Index i = 0; i < numPreservableLocals; i++) {
+    for (Index i = 0; i < numLocals; i++) {
+      if (!relevantLiveLocals.count(i)) {
+        continue;
+      }
       auto localType = func->getLocalType(i);
       const auto& types = localType.expand();
       for (Index j = 0; j < types.size(); j++) {
diff --git a/test/passes/asyncify.txt b/test/passes/asyncify.txt
new file mode 100644
index 000000000..4137c946c
--- /dev/null
+++ b/test/passes/asyncify.txt
@@ -0,0 +1,1261 @@
+(module
+ (type $i32_=>_none (func (param i32)))
+ (type $i32_i32_=>_none (func (param i32 i32)))
+ (type $none_=>_none (func))
+ (type $none_=>_i32 (func (result i32)))
+ (import "env" "import" (func $import))
+ (import "env" "import2" (func $import2 (param i32)))
+ (memory $0 1 2)
+ (table $0 2 2 funcref)
+ (elem (i32.const 0) $liveness2 $liveness2)
+ (global $__asyncify_state (mut i32) (i32.const 0))
+ (global $__asyncify_data (mut i32) (i32.const 0))
+ (export "asyncify_start_unwind" (func $asyncify_start_unwind))
+ (export "asyncify_stop_unwind" (func $asyncify_stop_unwind))
+ (export "asyncify_start_rewind" (func $asyncify_start_rewind))
+ (export "asyncify_stop_rewind" (func $asyncify_stop_rewind))
+ (export "asyncify_get_state" (func $asyncify_get_state))
+ (func $liveness1 (param $live0 i32) (param $dead0 i32)
+  (local $live1 i32)
+  (local $dead1 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (local $7 i32)
+  (local $8 i32)
+  (local $9 i32)
+  (local $10 i32)
+  (local $11 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -8)
+     )
+    )
+    (local.set $10
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $live0
+     (i32.load
+      (local.get $10)
+     )
+    )
+    (local.set $live1
+     (i32.load offset=4
+      (local.get $10)
+     )
+    )
+   )
+  )
+  (local.set $8
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $9
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $4
+          (local.get $dead0)
+         )
+         (drop
+          (local.get $4)
+         )
+         (local.set $5
+          (local.get $dead1)
+         )
+         (drop
+          (local.get $5)
+         )
+        )
+       )
+       (nop)
+       (nop)
+       (nop)
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $9)
+          (i32.const 0)
+         )
+        )
+        (block
+         (call $import)
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 0)
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $6
+          (local.get $live0)
+         )
+         (drop
+          (local.get $6)
+         )
+         (local.set $7
+          (local.get $live1)
+         )
+         (drop
+          (local.get $7)
+         )
+        )
+       )
+       (nop)
+       (nop)
+       (nop)
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $8)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $11
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $11)
+    (local.get $live0)
+   )
+   (i32.store offset=4
+    (local.get $11)
+    (local.get $live1)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 8)
+    )
+   )
+  )
+ )
+ (func $liveness2 (param $live0 i32) (param $dead0 i32)
+  (local $live1 i32)
+  (local $dead1 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (local $7 i32)
+  (local $8 i32)
+  (local $9 i32)
+  (local $10 i32)
+  (local $11 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -8)
+     )
+    )
+    (local.set $10
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $live0
+     (i32.load
+      (local.get $10)
+     )
+    )
+    (local.set $live1
+     (i32.load offset=4
+      (local.get $10)
+     )
+    )
+   )
+  )
+  (local.set $8
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $9
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $4
+          (local.get $dead0)
+         )
+         (drop
+          (local.get $4)
+         )
+         (local.set $5
+          (local.get $dead1)
+         )
+         (drop
+          (local.get $5)
+         )
+        )
+       )
+       (nop)
+       (nop)
+       (nop)
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $9)
+          (i32.const 0)
+         )
+        )
+        (block
+         (call $import)
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 0)
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $6
+          (local.get $live0)
+         )
+         (drop
+          (local.get $6)
+         )
+         (local.set $7
+          (local.get $live1)
+         )
+         (drop
+          (local.get $7)
+         )
+        )
+       )
+       (nop)
+       (nop)
+       (nop)
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $8)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $11
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $11)
+    (local.get $live0)
+   )
+   (i32.store offset=4
+    (local.get $11)
+    (local.get $live1)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 8)
+    )
+   )
+  )
+ )
+ (func $liveness3 (param $live0 i32) (param $dead0 i32)
+  (local $live1 i32)
+  (local $dead1 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (local $7 i32)
+  (local $8 i32)
+  (local $9 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -8)
+     )
+    )
+    (local.set $8
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $live0
+     (i32.load
+      (local.get $8)
+     )
+    )
+    (local.set $live1
+     (i32.load offset=4
+      (local.get $8)
+     )
+    )
+   )
+  )
+  (local.set $6
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $7
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $7)
+          (i32.const 0)
+         )
+        )
+        (block
+         (call $import)
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 0)
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $4
+          (local.get $live0)
+         )
+         (drop
+          (local.get $4)
+         )
+        )
+       )
+       (nop)
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $7)
+          (i32.const 1)
+         )
+        )
+        (block
+         (call $import)
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 1)
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $5
+          (local.get $live1)
+         )
+         (drop
+          (local.get $5)
+         )
+        )
+       )
+       (nop)
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $6)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $9
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $9)
+    (local.get $live0)
+   )
+   (i32.store offset=4
+    (local.get $9)
+    (local.get $live1)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 8)
+    )
+   )
+  )
+ )
+ (func $liveness4 (param $live0 i32) (param $dead0 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -4)
+     )
+    )
+    (local.set $5
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $live0
+     (i32.load
+      (local.get $5)
+     )
+    )
+   )
+  )
+  (local.set $3
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $4
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.or
+         (i32.const 0)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 2)
+         )
+        )
+        (if
+         (if (result i32)
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 0)
+          )
+          (i32.const 1)
+          (i32.eq
+           (local.get $4)
+           (i32.const 0)
+          )
+         )
+         (block
+          (call $import)
+          (if
+           (i32.eq
+            (global.get $__asyncify_state)
+            (i32.const 1)
+           )
+           (br $__asyncify_unwind
+            (i32.const 0)
+           )
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $2
+          (local.get $live0)
+         )
+         (drop
+          (local.get $2)
+         )
+        )
+       )
+       (nop)
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $3)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $6
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $6)
+    (local.get $live0)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+ )
+ (func $liveness5 (param $dead0 i32)
+  (local $1 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (nop)
+  )
+  (local.set $3
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $4
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $1
+          (local.get $dead0)
+         )
+         (drop
+          (local.get $1)
+         )
+        )
+       )
+       (nop)
+       (if
+        (i32.or
+         (i32.const 0)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 2)
+         )
+        )
+        (if
+         (if (result i32)
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 0)
+          )
+          (i32.const 1)
+          (i32.eq
+           (local.get $4)
+           (i32.const 0)
+          )
+         )
+         (block
+          (call $import)
+          (if
+           (i32.eq
+            (global.get $__asyncify_state)
+            (i32.const 1)
+           )
+           (br $__asyncify_unwind
+            (i32.const 0)
+           )
+          )
+         )
+        )
+       )
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $dead0
+          (i32.const 1)
+         )
+         (local.set $2
+          (local.get $dead0)
+         )
+         (drop
+          (local.get $2)
+         )
+        )
+       )
+       (nop)
+       (nop)
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $3)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (nop)
+ )
+ (func $liveness-call-kills (param $live i32)
+  (local $1 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -4)
+     )
+    )
+    (local.set $4
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $1
+     (i32.load
+      (local.get $4)
+     )
+    )
+   )
+  )
+  (local.set $2
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $3
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (local.set $1
+         (local.get $live)
+        )
+       )
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $3)
+          (i32.const 0)
+         )
+        )
+        (block
+         (call $import2
+          (local.get $1)
+         )
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 0)
+          )
+         )
+        )
+       )
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $2)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $5
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $5)
+    (local.get $1)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+ )
+ (func $liveness-indirect-kills (param $live0 i32) (param $live1 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (local $7 i32)
+  (if
+   (i32.eq
+    (global.get $__asyncify_state)
+    (i32.const 2)
+   )
+   (block
+    (i32.store
+     (global.get $__asyncify_data)
+     (i32.add
+      (i32.load
+       (global.get $__asyncify_data)
+      )
+      (i32.const -8)
+     )
+    )
+    (local.set $6
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+    )
+    (local.set $2
+     (i32.load
+      (local.get $6)
+     )
+    )
+    (local.set $3
+     (i32.load offset=4
+      (local.get $6)
+     )
+    )
+   )
+  )
+  (local.set $4
+   (block $__asyncify_unwind (result i32)
+    (block
+     (block
+      (if
+       (i32.eq
+        (global.get $__asyncify_state)
+        (i32.const 2)
+       )
+       (block
+        (i32.store
+         (global.get $__asyncify_data)
+         (i32.add
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+          (i32.const -4)
+         )
+        )
+        (local.set $5
+         (i32.load
+          (i32.load
+           (global.get $__asyncify_data)
+          )
+         )
+        )
+       )
+      )
+      (block
+       (if
+        (i32.eq
+         (global.get $__asyncify_state)
+         (i32.const 0)
+        )
+        (block
+         (local.set $2
+          (local.get $live0)
+         )
+         (local.set $3
+          (local.get $live1)
+         )
+        )
+       )
+       (nop)
+       (if
+        (if (result i32)
+         (i32.eq
+          (global.get $__asyncify_state)
+          (i32.const 0)
+         )
+         (i32.const 1)
+         (i32.eq
+          (local.get $5)
+          (i32.const 0)
+         )
+        )
+        (block
+         (call_indirect (type $i32_=>_none)
+          (local.get $2)
+          (local.get $3)
+         )
+         (if
+          (i32.eq
+           (global.get $__asyncify_state)
+           (i32.const 1)
+          )
+          (br $__asyncify_unwind
+           (i32.const 0)
+          )
+         )
+        )
+       )
+      )
+     )
+     (return)
+    )
+   )
+  )
+  (block
+   (i32.store
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (local.get $4)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (local.set $7
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+   )
+   (i32.store
+    (local.get $7)
+    (local.get $2)
+   )
+   (i32.store offset=4
+    (local.get $7)
+    (local.get $3)
+   )
+   (i32.store
+    (global.get $__asyncify_data)
+    (i32.add
+     (i32.load
+      (global.get $__asyncify_data)
+     )
+     (i32.const 8)
+    )
+   )
+  )
+ )
+ (func $asyncify_start_unwind (param $0 i32)
+  (global.set $__asyncify_state
+   (i32.const 1)
+  )
+  (global.set $__asyncify_data
+   (local.get $0)
+  )
+  (if
+   (i32.gt_u
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (i32.load offset=4
+     (global.get $__asyncify_data)
+    )
+   )
+   (unreachable)
+  )
+ )
+ (func $asyncify_stop_unwind
+  (global.set $__asyncify_state
+   (i32.const 0)
+  )
+  (if
+   (i32.gt_u
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (i32.load offset=4
+     (global.get $__asyncify_data)
+    )
+   )
+   (unreachable)
+  )
+ )
+ (func $asyncify_start_rewind (param $0 i32)
+  (global.set $__asyncify_state
+   (i32.const 2)
+  )
+  (global.set $__asyncify_data
+   (local.get $0)
+  )
+  (if
+   (i32.gt_u
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (i32.load offset=4
+     (global.get $__asyncify_data)
+    )
+   )
+   (unreachable)
+  )
+ )
+ (func $asyncify_stop_rewind
+  (global.set $__asyncify_state
+   (i32.const 0)
+  )
+  (if
+   (i32.gt_u
+    (i32.load
+     (global.get $__asyncify_data)
+    )
+    (i32.load offset=4
+     (global.get $__asyncify_data)
+    )
+   )
+   (unreachable)
+  )
+ )
+ (func $asyncify_get_state (result i32)
+  (global.get $__asyncify_state)
+ )
+)
diff --git a/test/passes/asyncify.wast b/test/passes/asyncify.wast
new file mode 100644
index 000000000..281934aea
--- /dev/null
+++ b/test/passes/asyncify.wast
@@ -0,0 +1,53 @@
+(module
+  (type $f (func (param i32)))
+  (memory 1 2)
+  (import "env" "import" (func $import))
+  (import "env" "import2" (func $import2 (param i32)))
+  (table 1 1)
+  (func $liveness1 (param $live0 i32) (param $dead0 i32)
+    (local $live1 i32)
+    (local $dead1 i32)
+    (drop (local.get $dead0))
+    (drop (local.get $dead1))
+    (call $import)
+    (drop (local.get $live0))
+    (drop (local.get $live1))
+  )
+  (func $liveness2 (param $live0 i32) (param $dead0 i32)
+    (local $live1 i32)
+    (local $dead1 i32)
+    (drop (local.get $dead0))
+    (drop (local.get $dead1))
+    (call $import)
+    (drop (local.get $live0))
+    (drop (local.get $live1))
+  )
+  (func $liveness3 (param $live0 i32) (param $dead0 i32)
+    (local $live1 i32)
+    (local $dead1 i32)
+    (call $import)
+    (drop (local.get $live0))
+    (call $import)
+    (drop (local.get $live1))
+  )
+  (func $liveness4 (param $live0 i32) (param $dead0 i32)
+    (if (i32.const 0)
+      (call $import)
+    )
+    (drop (local.get $live0))
+  )
+  (func $liveness5 (param $dead0 i32)
+    (drop (local.get $dead0))
+    (if (i32.const 0)
+      (call $import) ;; live before and after call, but not during
+    )
+    (local.set $dead0 (i32.const 1))
+    (drop (local.get $dead0))
+  )
+  (func $liveness-call-kills (param $live i32)
+    (call $import2 (local.get $live))
+  )
+  (func $liveness-indirect-kills (param $live0 i32) (param $live1 i32)
+    (call_indirect (type $f) (local.get $live0) (local.get $live1))
+  )
+)
diff --git a/test/passes/asyncify_enable-multivalue.txt b/test/passes/asyncify_enable-multivalue.txt
index 5aee73c7a..08d0d7bf9 100644
--- a/test/passes/asyncify_enable-multivalue.txt
+++ b/test/passes/asyncify_enable-multivalue.txt
@@ -447,7 +447,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -20)
+      (i32.const -4)
      )
     )
     (local.set $8
@@ -455,28 +455,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $temp
-     (i32.load
-      (local.get $8)
-     )
-    )
     (local.set $1
-     (i32.load offset=4
-      (local.get $8)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $8)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $8)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
+     (i32.load
       (local.get $8)
      )
     )
@@ -592,31 +572,15 @@
    )
    (i32.store
     (local.get $9)
-    (local.get $temp)
-   )
-   (i32.store offset=4
-    (local.get $9)
     (local.get $1)
    )
-   (i32.store offset=8
-    (local.get $9)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $9)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $9)
-    (local.get $4)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 20)
+     (i32.const 4)
     )
    )
   )
@@ -806,7 +770,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -52)
+      (i32.const -4)
      )
     )
     (local.set $13
@@ -814,63 +778,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $13)
-     )
-    )
     (local.set $y
-     (i32.load offset=4
-      (local.get $13)
-     )
-    )
-    (local.set $z
-     (tuple.make
-      (f32.load offset=8
-       (local.get $13)
-      )
-      (i64.load offset=12 align=4
-       (local.get $13)
-      )
-     )
-    )
-    (local.set $3
-     (i32.load offset=20
-      (local.get $13)
-     )
-    )
-    (local.set $4
-     (i32.load offset=24
-      (local.get $13)
-     )
-    )
-    (local.set $5
-     (i32.load offset=28
-      (local.get $13)
-     )
-    )
-    (local.set $6
-     (i32.load offset=32
-      (local.get $13)
-     )
-    )
-    (local.set $7
-     (i32.load offset=36
-      (local.get $13)
-     )
-    )
-    (local.set $8
-     (i32.load offset=40
-      (local.get $13)
-     )
-    )
-    (local.set $9
-     (i32.load offset=44
-      (local.get $13)
-     )
-    )
-    (local.set $10
-     (i32.load offset=48
+     (i32.load
       (local.get $13)
      )
     )
@@ -1015,63 +924,15 @@
    )
    (i32.store
     (local.get $14)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $14)
     (local.get $y)
    )
-   (f32.store offset=8
-    (local.get $14)
-    (tuple.extract 0
-     (local.get $z)
-    )
-   )
-   (i64.store offset=12 align=4
-    (local.get $14)
-    (tuple.extract 1
-     (local.get $z)
-    )
-   )
-   (i32.store offset=20
-    (local.get $14)
-    (local.get $3)
-   )
-   (i32.store offset=24
-    (local.get $14)
-    (local.get $4)
-   )
-   (i32.store offset=28
-    (local.get $14)
-    (local.get $5)
-   )
-   (i32.store offset=32
-    (local.get $14)
-    (local.get $6)
-   )
-   (i32.store offset=36
-    (local.get $14)
-    (local.get $7)
-   )
-   (i32.store offset=40
-    (local.get $14)
-    (local.get $8)
-   )
-   (i32.store offset=44
-    (local.get $14)
-    (local.get $9)
-   )
-   (i32.store offset=48
-    (local.get $14)
-    (local.get $10)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 52)
+     (i32.const 4)
     )
    )
   )
@@ -1081,39 +942,12 @@
   (local $1 i32)
   (local $2 i32)
   (local $3 i32)
-  (local $4 i32)
-  (local $5 i32)
   (if
    (i32.eq
     (global.get $__asyncify_state)
     (i32.const 2)
    )
-   (block
-    (i32.store
-     (global.get $__asyncify_data)
-     (i32.add
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-      (i32.const -8)
-     )
-    )
-    (local.set $4
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-    )
-    (local.set $x
-     (i32.load
-      (local.get $4)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $4)
-     )
-    )
-   )
+   (nop)
   )
   (local.set $2
    (block $__asyncify_unwind (result i32)
@@ -1210,30 +1044,7 @@
     )
    )
   )
-  (block
-   (local.set $5
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-   )
-   (i32.store
-    (local.get $5)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $5)
-    (local.get $1)
-   )
-   (i32.store
-    (global.get $__asyncify_data)
-    (i32.add
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-     (i32.const 8)
-    )
-   )
-  )
+  (nop)
  )
  (func $calls-import2-if-else (param $x i32)
   (local $1 i32)
@@ -1254,7 +1065,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -12)
+      (i32.const -4)
      )
     )
     (local.set $5
@@ -1262,18 +1073,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $5)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $5)
-     )
-    )
     (local.set $2
-     (i32.load offset=8
+     (i32.load
       (local.get $5)
      )
     )
@@ -1433,14 +1234,6 @@
    )
    (i32.store
     (local.get $6)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $6)
-    (local.get $1)
-   )
-   (i32.store offset=8
-    (local.get $6)
     (local.get $2)
    )
    (i32.store
@@ -1449,7 +1242,7 @@
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 12)
+     (i32.const 4)
     )
    )
   )
@@ -1461,54 +1254,12 @@
   (local $4 i32)
   (local $5 i32)
   (local $6 i32)
-  (local $7 i32)
-  (local $8 i32)
   (if
    (i32.eq
     (global.get $__asyncify_state)
     (i32.const 2)
    )
-   (block
-    (i32.store
-     (global.get $__asyncify_data)
-     (i32.add
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-      (i32.const -20)
-     )
-    )
-    (local.set $7
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-    )
-    (local.set $x
-     (i32.load
-      (local.get $7)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $7)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $7)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $7)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
-      (local.get $7)
-     )
-    )
-   )
+   (nop)
   )
   (local.set $5
    (block $__asyncify_unwind (result i32)
@@ -1650,42 +1401,7 @@
     )
    )
   )
-  (block
-   (local.set $8
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-   )
-   (i32.store
-    (local.get $8)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $8)
-    (local.get $1)
-   )
-   (i32.store offset=8
-    (local.get $8)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $8)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $8)
-    (local.get $4)
-   )
-   (i32.store
-    (global.get $__asyncify_data)
-    (i32.add
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-     (i32.const 20)
-    )
-   )
-  )
+  (nop)
   (i32.const 0)
  )
  (func $calls-import2-if-else-oneside2 (param $x i32) (result i32)
@@ -1709,7 +1425,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -20)
+      (i32.const -4)
      )
     )
     (local.set $7
@@ -1717,28 +1433,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $7)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $7)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $7)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $7)
-     )
-    )
     (local.set $4
-     (i32.load offset=16
+     (i32.load
       (local.get $7)
      )
     )
@@ -1892,22 +1588,6 @@
    )
    (i32.store
     (local.get $8)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $8)
-    (local.get $1)
-   )
-   (i32.store offset=8
-    (local.get $8)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $8)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $8)
     (local.get $4)
    )
    (i32.store
@@ -1916,7 +1596,7 @@
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 20)
+     (i32.const 4)
     )
    )
   )
@@ -1942,7 +1622,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -24)
+      (i32.const -12)
      )
     )
     (local.set $5
@@ -1950,7 +1630,7 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
+    (local.set $1
      (tuple.make
       (i32.load
        (local.get $5)
@@ -1960,16 +1640,6 @@
       )
      )
     )
-    (local.set $1
-     (tuple.make
-      (i32.load offset=12
-       (local.get $5)
-      )
-      (i64.load offset=16 align=4
-       (local.get $5)
-      )
-     )
-    )
    )
   )
   (local.set $2
@@ -2072,22 +1742,10 @@
    (i32.store
     (local.get $6)
     (tuple.extract 0
-     (local.get $x)
-    )
-   )
-   (i64.store offset=4 align=4
-    (local.get $6)
-    (tuple.extract 1
-     (local.get $x)
-    )
-   )
-   (i32.store offset=12
-    (local.get $6)
-    (tuple.extract 0
      (local.get $1)
     )
    )
-   (i64.store offset=16 align=4
+   (i64.store offset=4 align=4
     (local.get $6)
     (tuple.extract 1
      (local.get $1)
@@ -2099,7 +1757,7 @@
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 24)
+     (i32.const 12)
     )
    )
   )
@@ -2124,7 +1782,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -16)
+      (i32.const -4)
      )
     )
     (local.set $6
@@ -2137,21 +1795,6 @@
       (local.get $6)
      )
     )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $6)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $6)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $6)
-     )
-    )
    )
   )
   (local.set $4
@@ -2273,25 +1916,13 @@
     (local.get $7)
     (local.get $x)
    )
-   (i32.store offset=4
-    (local.get $7)
-    (local.get $1)
-   )
-   (i32.store offset=8
-    (local.get $7)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $7)
-    (local.get $3)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 16)
+     (i32.const 4)
     )
    )
   )
diff --git a/test/passes/asyncify_mod-asyncify-always-and-only-unwind.txt b/test/passes/asyncify_mod-asyncify-always-and-only-unwind.txt
index 7be0eed18..62c3ba33f 100644
--- a/test/passes/asyncify_mod-asyncify-always-and-only-unwind.txt
+++ b/test/passes/asyncify_mod-asyncify-always-and-only-unwind.txt
@@ -111,7 +111,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -20)
+      (i32.const -4)
      )
     )
     (local.set $8
@@ -119,28 +119,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $temp
-     (i32.load
-      (local.get $8)
-     )
-    )
     (local.set $1
-     (i32.load offset=4
-      (local.get $8)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $8)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $8)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
+     (i32.load
       (local.get $8)
      )
     )
@@ -250,31 +230,15 @@
    )
    (i32.store
     (local.get $9)
-    (local.get $temp)
-   )
-   (i32.store offset=4
-    (local.get $9)
     (local.get $1)
    )
-   (i32.store offset=8
-    (local.get $9)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $9)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $9)
-    (local.get $4)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 20)
+     (i32.const 4)
     )
    )
   )
diff --git a/test/passes/asyncify_mod-asyncify-never-unwind.txt b/test/passes/asyncify_mod-asyncify-never-unwind.txt
index 264760759..eaf2794cf 100644
--- a/test/passes/asyncify_mod-asyncify-never-unwind.txt
+++ b/test/passes/asyncify_mod-asyncify-never-unwind.txt
@@ -120,7 +120,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -20)
+      (i32.const -4)
      )
     )
     (local.set $8
@@ -128,28 +128,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $temp
-     (i32.load
-      (local.get $8)
-     )
-    )
     (local.set $1
-     (i32.load offset=4
-      (local.get $8)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $8)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $8)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
+     (i32.load
       (local.get $8)
      )
     )
@@ -262,31 +242,15 @@
    )
    (i32.store
     (local.get $9)
-    (local.get $temp)
-   )
-   (i32.store offset=4
-    (local.get $9)
     (local.get $1)
    )
-   (i32.store offset=8
-    (local.get $9)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $9)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $9)
-    (local.get $4)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 20)
+     (i32.const 4)
     )
    )
   )
diff --git a/test/passes/asyncify_optimize-level=1.txt b/test/passes/asyncify_optimize-level=1.txt
index 1946a7f96..19e0cd160 100644
--- a/test/passes/asyncify_optimize-level=1.txt
+++ b/test/passes/asyncify_optimize-level=1.txt
@@ -286,26 +286,19 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -8)
+      (i32.const -4)
      )
     )
-    (local.set $0
+    (local.set $1
      (i32.load
-      (local.tee $1
-       (i32.load
-        (global.get $__asyncify_data)
-       )
+      (i32.load
+       (global.get $__asyncify_data)
       )
      )
     )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $1)
-     )
-    )
    )
   )
-  (local.set $2
+  (local.set $0
    (block $__asyncify_unwind (result i32)
     (if
      (i32.eq
@@ -385,7 +378,7 @@
    (i32.load
     (global.get $__asyncify_data)
    )
-   (local.get $2)
+   (local.get $0)
   )
   (i32.store
    (global.get $__asyncify_data)
@@ -397,15 +390,9 @@
    )
   )
   (i32.store
-   (local.tee $2
-    (i32.load
-     (global.get $__asyncify_data)
-    )
+   (i32.load
+    (global.get $__asyncify_data)
    )
-   (local.get $0)
-  )
-  (i32.store offset=4
-   (local.get $2)
    (local.get $1)
   )
   (i32.store
@@ -414,38 +401,14 @@
     (i32.load
      (global.get $__asyncify_data)
     )
-    (i32.const 8)
+    (i32.const 4)
    )
   )
   (i32.const 0)
  )
  (func $calls-import2-if (param $0 i32)
   (local $1 i32)
-  (if
-   (i32.eq
-    (global.get $__asyncify_state)
-    (i32.const 2)
-   )
-   (block
-    (i32.store
-     (global.get $__asyncify_data)
-     (i32.add
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-      (i32.const -4)
-     )
-    )
-    (local.set $0
-     (i32.load
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-     )
-    )
-   )
-  )
-  (local.set $1
+  (local.set $0
    (block $__asyncify_unwind (result i32)
     (if
      (i32.eq
@@ -508,21 +471,6 @@
    (i32.load
     (global.get $__asyncify_data)
    )
-   (local.get $1)
-  )
-  (i32.store
-   (global.get $__asyncify_data)
-   (i32.add
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-    (i32.const 4)
-   )
-  )
-  (i32.store
-   (i32.load
-    (global.get $__asyncify_data)
-   )
    (local.get $0)
   )
   (i32.store
@@ -550,26 +498,19 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -8)
+      (i32.const -4)
      )
     )
-    (local.set $0
+    (local.set $1
      (i32.load
-      (local.tee $1
-       (i32.load
-        (global.get $__asyncify_data)
-       )
+      (i32.load
+       (global.get $__asyncify_data)
       )
      )
     )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $1)
-     )
-    )
    )
   )
-  (local.set $2
+  (local.set $0
    (block $__asyncify_unwind (result i32)
     (if
      (i32.eq
@@ -675,7 +616,7 @@
    (i32.load
     (global.get $__asyncify_data)
    )
-   (local.get $2)
+   (local.get $0)
   )
   (i32.store
    (global.get $__asyncify_data)
@@ -687,15 +628,9 @@
    )
   )
   (i32.store
-   (local.tee $2
-    (i32.load
-     (global.get $__asyncify_data)
-    )
+   (i32.load
+    (global.get $__asyncify_data)
    )
-   (local.get $0)
-  )
-  (i32.store offset=4
-   (local.get $2)
    (local.get $1)
   )
   (i32.store
@@ -704,45 +639,14 @@
     (i32.load
      (global.get $__asyncify_data)
     )
-    (i32.const 8)
+    (i32.const 4)
    )
   )
  )
  (func $calls-import2-if-else-oneside (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
-  (if
-   (i32.eq
-    (global.get $__asyncify_state)
-    (i32.const 2)
-   )
-   (block
-    (i32.store
-     (global.get $__asyncify_data)
-     (i32.add
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-      (i32.const -8)
-     )
-    )
-    (local.set $0
-     (i32.load
-      (local.tee $1
-       (i32.load
-        (global.get $__asyncify_data)
-       )
-      )
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $1)
-     )
-    )
-   )
-  )
-  (local.set $2
+  (local.set $0
    (block $__asyncify_unwind (result i32)
     (if
      (i32.eq
@@ -759,7 +663,7 @@
         (i32.const -4)
        )
       )
-      (local.set $2
+      (local.set $1
        (i32.load
         (i32.load
          (global.get $__asyncify_data)
@@ -770,9 +674,9 @@
     )
     (if
      (i32.or
-      (local.tee $1
+      (local.tee $2
        (select
-        (local.get $1)
+        (local.get $2)
         (local.get $0)
         (global.get $__asyncify_state)
        )
@@ -794,7 +698,7 @@
     (if
      (i32.or
       (i32.eqz
-       (local.get $1)
+       (local.get $2)
       )
       (i32.eq
        (global.get $__asyncify_state)
@@ -804,7 +708,7 @@
      (if
       (select
        (i32.eqz
-        (local.get $2)
+        (local.get $1)
        )
        (i32.const 1)
        (global.get $__asyncify_state)
@@ -840,36 +744,15 @@
    (i32.load
     (global.get $__asyncify_data)
    )
-   (local.get $2)
-  )
-  (i32.store
-   (global.get $__asyncify_data)
-   (i32.add
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-    (i32.const 4)
-   )
-  )
-  (i32.store
-   (local.tee $2
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-   )
    (local.get $0)
   )
-  (i32.store offset=4
-   (local.get $2)
-   (local.get $1)
-  )
   (i32.store
    (global.get $__asyncify_data)
    (i32.add
     (i32.load
      (global.get $__asyncify_data)
     )
-    (i32.const 8)
+    (i32.const 4)
    )
   )
   (i32.const 0)
@@ -889,26 +772,19 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -8)
+      (i32.const -4)
      )
     )
-    (local.set $0
+    (local.set $1
      (i32.load
-      (local.tee $1
-       (i32.load
-        (global.get $__asyncify_data)
-       )
+      (i32.load
+       (global.get $__asyncify_data)
       )
      )
     )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $1)
-     )
-    )
    )
   )
-  (local.set $2
+  (local.set $0
    (block $__asyncify_unwind (result i32)
     (if
      (i32.eq
@@ -1006,7 +882,7 @@
    (i32.load
     (global.get $__asyncify_data)
    )
-   (local.get $2)
+   (local.get $0)
   )
   (i32.store
    (global.get $__asyncify_data)
@@ -1018,15 +894,9 @@
    )
   )
   (i32.store
-   (local.tee $2
-    (i32.load
-     (global.get $__asyncify_data)
-    )
+   (i32.load
+    (global.get $__asyncify_data)
    )
-   (local.get $0)
-  )
-  (i32.store offset=4
-   (local.get $2)
    (local.get $1)
   )
   (i32.store
@@ -1035,7 +905,7 @@
     (i32.load
      (global.get $__asyncify_data)
     )
-    (i32.const 8)
+    (i32.const 4)
    )
   )
   (i32.const 0)
diff --git a/test/passes/asyncify_pass-arg=asyncify-ignore-imports.txt b/test/passes/asyncify_pass-arg=asyncify-ignore-imports.txt
index b1854fb5d..f9ae663fd 100644
--- a/test/passes/asyncify_pass-arg=asyncify-ignore-imports.txt
+++ b/test/passes/asyncify_pass-arg=asyncify-ignore-imports.txt
@@ -60,7 +60,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -8)
+      (i32.const -4)
      )
     )
     (local.set $4
@@ -68,13 +68,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $4)
-     )
-    )
     (local.set $1
-     (i32.load offset=4
+     (i32.load
       (local.get $4)
      )
     )
@@ -176,10 +171,6 @@
    )
    (i32.store
     (local.get $5)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $5)
     (local.get $1)
    )
    (i32.store
@@ -188,7 +179,7 @@
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 8)
+     (i32.const 4)
     )
    )
   )
diff --git a/test/passes/asyncify_pass-arg=asyncify-ignore-indirect.txt b/test/passes/asyncify_pass-arg=asyncify-ignore-indirect.txt
index a48f3b3eb..690c96030 100644
--- a/test/passes/asyncify_pass-arg=asyncify-ignore-indirect.txt
+++ b/test/passes/asyncify_pass-arg=asyncify-ignore-indirect.txt
@@ -267,7 +267,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -12)
+      (i32.const -4)
      )
     )
     (local.set $5
@@ -275,18 +275,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $5)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $5)
-     )
-    )
     (local.set $2
-     (i32.load offset=8
+     (i32.load
       (local.get $5)
      )
     )
@@ -446,14 +436,6 @@
    )
    (i32.store
     (local.get $6)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $6)
-    (local.get $1)
-   )
-   (i32.store offset=8
-    (local.get $6)
     (local.get $2)
    )
    (i32.store
@@ -462,7 +444,7 @@
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 12)
+     (i32.const 4)
     )
    )
   )
diff --git a/test/passes/asyncify_pass-arg=asyncify-imports@env.import,env.import2.txt b/test/passes/asyncify_pass-arg=asyncify-imports@env.import,env.import2.txt
index 48187bef3..defef94ed 100644
--- a/test/passes/asyncify_pass-arg=asyncify-imports@env.import,env.import2.txt
+++ b/test/passes/asyncify_pass-arg=asyncify-imports@env.import,env.import2.txt
@@ -444,7 +444,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -20)
+      (i32.const -4)
      )
     )
     (local.set $8
@@ -452,28 +452,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $temp
-     (i32.load
-      (local.get $8)
-     )
-    )
     (local.set $1
-     (i32.load offset=4
-      (local.get $8)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $8)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $8)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
+     (i32.load
       (local.get $8)
      )
     )
@@ -589,31 +569,15 @@
    )
    (i32.store
     (local.get $9)
-    (local.get $temp)
-   )
-   (i32.store offset=4
-    (local.get $9)
     (local.get $1)
    )
-   (i32.store offset=8
-    (local.get $9)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $9)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $9)
-    (local.get $4)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 20)
+     (i32.const 4)
     )
    )
   )
@@ -802,7 +766,7 @@
       (i32.load
        (global.get $__asyncify_data)
       )
-      (i32.const -40)
+      (i32.const -4)
      )
     )
     (local.set $12
@@ -810,53 +774,8 @@
       (global.get $__asyncify_data)
      )
     )
-    (local.set $x
-     (i32.load
-      (local.get $12)
-     )
-    )
     (local.set $y
-     (i32.load offset=4
-      (local.get $12)
-     )
-    )
-    (local.set $2
-     (i32.load offset=8
-      (local.get $12)
-     )
-    )
-    (local.set $3
-     (i32.load offset=12
-      (local.get $12)
-     )
-    )
-    (local.set $4
-     (i32.load offset=16
-      (local.get $12)
-     )
-    )
-    (local.set $5
-     (i32.load offset=20
-      (local.get $12)
-     )
-    )
-    (local.set $6
-     (i32.load offset=24
-      (local.get $12)
-     )
-    )
-    (local.set $7
-     (i32.load offset=28
-      (local.get $12)
-     )
-    )
-    (local.set $8
-     (i32.load offset=32
-      (local.get $12)
-     )
-    )
-    (local.set $9
-     (i32.load offset=36
+     (i32.load
       (local.get $12)
      )
     )
@@ -1001,51 +920,15 @@
    )
    (i32.store
     (local.get $13)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $13)
     (local.get $y)
    )
-   (i32.store offset=8
-    (local.get $13)
-    (local.get $2)
-   )
-   (i32.store offset=12
-    (local.get $13)
-    (local.get $3)
-   )
-   (i32.store offset=16
-    (local.get $13)
-    (local.get $4)
-   )
-   (i32.store offset=20
-    (local.get $13)
-    (local.get $5)
-   )
-   (i32.store offset=24
-    (local.get $13)
-    (local.get $6)
-   )
-   (i32.store offset=28
-    (local.get $13)
-    (local.get $7)
-   )
-   (i32.store offset=32
-    (local.get $13)
-    (local.get $8)
-   )
-   (i32.store offset=36
-    (local.get $13)
-    (local.get $9)
-   )
    (i32.store
     (global.get $__asyncify_data)
     (i32.add
      (i32.load
       (global.get $__asyncify_data)
      )
-     (i32.const 40)
+     (i32.const 4)
     )
    )
   )
@@ -1055,39 +938,12 @@
   (local $1 i32)
   (local $2 i32)
   (local $3 i32)
-  (local $4 i32)
-  (local $5 i32)
   (if
    (i32.eq
     (global.get $__asyncify_state)
     (i32.const 2)
    )
-   (block
-    (i32.store
-     (global.get $__asyncify_data)
-     (i32.add
-      (i32.load
-       (global.get $__asyncify_data)
-      )
-      (i32.const -8)
-     )
-    )
-    (local.set $4
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-    )
-    (local.set $x
-     (i32.load
-      (local.get $4)
-     )
-    )
-    (local.set $1
-     (i32.load offset=4
-      (local.get $4)
-     )
-    )
-   )
+   (nop)
   )
   (local.set $2
    (block $__asyncify_unwind (result i32)
@@ -1184,30 +1040,7 @@
     )
    )
   )
-  (block
-   (local.set $5
-    (i32.load
-     (global.get $__asyncify_data)
-    )
-   )
-   (i32.store
-    (local.get $5)
-    (local.get $x)
-   )
-   (i32.store offset=4
-    (local.get $5)
-    (local.get $1)
-   )
-   (i32.store
-    (global.get $__asyncify_data)
-    (i32.add
-     (i32.load
-      (global.get $__asyncify_data)
-     )
-     (i32.const 8)
-    )
-   )
-  )
+  (nop)
  )
  (func $calls-import2-if-else (param $x i32)
   (local $1 i32)