Fix regression from #850 (#851)

* fix regression from #850 - it is not always safe to fold added offsets into load/store offsets, as the add wraps but offset does not

* small refactoring to simplify asm2wasm pass invocation

10 files changed, 593 insertions, 378 deletions

diff --git a/src/asm2wasm.h b/src/asm2wasm.h
index 9cf8cf75a..2a6bcd0e7 100644
--- a/src/asm2wasm.h
+++ b/src/asm2wasm.h
@@ -859,13 +859,6 @@ void Asm2WasmBuilder::processAsm(Ref ast) {
 
   if (runOptimizationPasses) {
     optimizingBuilder->finish();
-    PassRunner passRunner(&wasm);
-    if (debug) {
-      passRunner.setDebug(true);
-      passRunner.setValidateGlobally(false);
-    }
-    passRunner.add("post-emscripten");
-    passRunner.run();
   }
 
   // second pass. first, function imports
@@ -993,6 +986,7 @@ void Asm2WasmBuilder::processAsm(Ref ast) {
     passRunner.add("vacuum");
     passRunner.add("remove-unused-brs");
     passRunner.add("optimize-instructions");
+    passRunner.add("post-emscripten");
   }
   passRunner.run();
 
diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp
index 37207ffc7..04216f1ee 100644
--- a/src/passes/OptimizeInstructions.cpp
+++ b/src/passes/OptimizeInstructions.cpp
@@ -436,33 +436,7 @@ private:
 
   // fold constant factors into the offset
   void optimizeMemoryAccess(Expression*& ptr, Address& offset) {
-    while (1) {
-      auto* add = ptr->dynCast<Binary>();
-      if (!add) break;
-      if (add->op != AddInt32) break;
-      auto* left = add->left->dynCast<Const>();
-      auto* right = add->right->dynCast<Const>();
-      // note: in optimized code, we shouldn't see an add of two constants, so don't worry about that much
-      // (precompute would optimize that)
-      if (left) {
-        auto value = left->value.geti32();
-        if (value >= 0) {
-          offset = offset + value;
-          ptr = add->right;
-          continue;
-        }
-      }
-      if (right) {
-        auto value = right->value.geti32();
-        if (value >= 0) {
-          offset = offset + value;
-          ptr = add->left;
-          continue;
-        }
-      }
-      break;
-    }
-    // finally, ptr may be a const, but it isn't worth folding that in (we still have a const); in fact,
+    // ptr may be a const, but it isn't worth folding that in (we still have a const); in fact,
     // it's better to do the opposite for gzip purposes as well as for readability.
     auto* last = ptr->dynCast<Const>();
     if (last) {
diff --git a/src/passes/PostEmscripten.cpp b/src/passes/PostEmscripten.cpp
index e46270c50..9a28efb1c 100644
--- a/src/passes/PostEmscripten.cpp
+++ b/src/passes/PostEmscripten.cpp
@@ -44,36 +44,49 @@ struct PostEmscripten : public WalkerPass<PostWalker<PostEmscripten, Visitor<Pos
   // it's 0 or an unused section of memory that was reserved for mapped globlas).
   // Thus it is ok to optimize such small constants into Load offsets.
 
-  template<typename T>
-  void visitMemoryOp(T *curr) {
-    if (curr->offset) return;
-    Expression* ptr = curr->ptr;
-    auto add = ptr->dynCast<Binary>();
-    if (!add || add->op != AddInt32) return;
-    assert(add->type == i32);
-    auto c = add->right->dynCast<Const>();
-    if (!c) {
-      c = add->left->dynCast<Const>();
-      if (c) {
-        // if one is a const, it's ok to swap
-        add->left = add->right;
-        add->right = c;
+  #define SAFE_MAX 1024
+
+  void optimizeMemoryAccess(Expression*& ptr, Address& offset) {
+    while (1) {
+      auto* add = ptr->dynCast<Binary>();
+      if (!add) break;
+      if (add->op != AddInt32) break;
+      auto* left = add->left->dynCast<Const>();
+      auto* right = add->right->dynCast<Const>();
+      // note: in optimized code, we shouldn't see an add of two constants, so don't worry about that much
+      // (precompute would optimize that)
+      if (left) {
+        auto value = left->value.geti32();
+        if (value >= 0 && value < SAFE_MAX) {
+          offset = offset + value;
+          ptr = add->right;
+          continue;
+        }
+      }
+      if (right) {
+        auto value = right->value.geti32();
+        if (value >= 0 && value < SAFE_MAX) {
+          offset = offset + value;
+          ptr = add->left;
+          continue;
+        }
       }
+      break;
     }
-    if (!c) return;
-    auto value = c->value.geti32();
-    if (value >= 0 && value < 1024) {
-      // foldable, by the above logic
-      curr->ptr = add->left;
-      curr->offset = value;
+    // finally ptr may be a const, but it isn't worth folding that in (we still have a const); in fact,
+    // it's better to do the opposite for gzip purposes as well as for readability.
+    auto* last = ptr->dynCast<Const>();
+    if (last) {
+      last->value = Literal(int32_t(last->value.geti32() + offset));
+      offset = 0;
     }
   }
 
   void visitLoad(Load* curr) {
-    visitMemoryOp(curr);
+    optimizeMemoryAccess(curr->ptr, curr->offset);
   }
   void visitStore(Store* curr) {
-    visitMemoryOp(curr);
+    optimizeMemoryAccess(curr->ptr, curr->offset);
   }
 };
 
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
index afb161718..06f88befe 100644
--- a/test/emcc_hello_world.fromasm
+++ b/test/emcc_hello_world.fromasm
@@ -4357,11 +4357,14 @@
                                           (i32.store8
                                             (get_local $5)
                                             (i32.or
-                                              (i32.load8_u offset=4075
-                                                (tee_local $7
-                                                  (call $f64-to-int
-                                                    (get_local $15)
+                                              (i32.load8_u
+                                                (i32.add
+                                                  (tee_local $7
+                                                    (call $f64-to-int
+                                                      (get_local $15)
+                                                    )
                                                   )
+                                                  (i32.const 4075)
                                                 )
                                               )
                                               (get_local $13)
@@ -6596,10 +6599,13 @@
                                   )
                                 )
                                 (i32.or
-                                  (i32.load8_u offset=4075
-                                    (i32.and
-                                      (get_local $5)
-                                      (i32.const 15)
+                                  (i32.load8_u
+                                    (i32.add
+                                      (i32.and
+                                        (get_local $5)
+                                        (i32.const 15)
+                                      )
+                                      (i32.const 4075)
                                     )
                                   )
                                   (get_local $9)
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
index adadfe086..4363dc349 100644
--- a/test/emcc_hello_world.fromasm.imprecise
+++ b/test/emcc_hello_world.fromasm.imprecise
@@ -4349,11 +4349,14 @@
                                           (i32.store8
                                             (get_local $5)
                                             (i32.or
-                                              (i32.load8_u offset=4075
-                                                (tee_local $7
-                                                  (i32.trunc_s/f64
-                                                    (get_local $15)
+                                              (i32.load8_u
+                                                (i32.add
+                                                  (tee_local $7
+                                                    (i32.trunc_s/f64
+                                                      (get_local $15)
+                                                    )
                                                   )
+                                                  (i32.const 4075)
                                                 )
                                               )
                                               (get_local $13)
@@ -6582,10 +6585,13 @@
                                   )
                                 )
                                 (i32.or
-                                  (i32.load8_u offset=4075
-                                    (i32.and
-                                      (get_local $5)
-                                      (i32.const 15)
+                                  (i32.load8_u
+                                    (i32.add
+                                      (i32.and
+                                        (get_local $5)
+                                        (i32.const 15)
+                                      )
+                                      (i32.const 4075)
                                     )
                                   )
                                   (get_local $9)
diff --git a/test/memorygrowth.fromasm b/test/memorygrowth.fromasm
index f61e92441..1422a65a9 100644
--- a/test/memorygrowth.fromasm
+++ b/test/memorygrowth.fromasm
@@ -723,80 +723,83 @@
                       (i32.and
                         (i32.load offset=4
                           (tee_local $14
-                            (i32.load offset=1512
-                              (i32.shl
-                                (i32.add
-                                  (i32.or
+                            (i32.load
+                              (i32.add
+                                (i32.shl
+                                  (i32.add
                                     (i32.or
                                       (i32.or
                                         (i32.or
-                                          (tee_local $8
+                                          (i32.or
+                                            (tee_local $8
+                                              (i32.and
+                                                (i32.shr_u
+                                                  (tee_local $4
+                                                    (i32.shr_u
+                                                      (get_local $8)
+                                                      (get_local $12)
+                                                    )
+                                                  )
+                                                  (i32.const 5)
+                                                )
+                                                (i32.const 8)
+                                              )
+                                            )
+                                            (get_local $12)
+                                          )
+                                          (tee_local $4
                                             (i32.and
                                               (i32.shr_u
-                                                (tee_local $4
+                                                (tee_local $1
                                                   (i32.shr_u
+                                                    (get_local $4)
                                                     (get_local $8)
-                                                    (get_local $12)
                                                   )
                                                 )
-                                                (i32.const 5)
+                                                (i32.const 2)
                                               )
-                                              (i32.const 8)
+                                              (i32.const 4)
                                             )
                                           )
-                                          (get_local $12)
                                         )
-                                        (tee_local $4
+                                        (tee_local $1
                                           (i32.and
                                             (i32.shr_u
-                                              (tee_local $1
+                                              (tee_local $0
                                                 (i32.shr_u
+                                                  (get_local $1)
                                                   (get_local $4)
-                                                  (get_local $8)
                                                 )
                                               )
-                                              (i32.const 2)
+                                              (i32.const 1)
                                             )
-                                            (i32.const 4)
+                                            (i32.const 2)
                                           )
                                         )
                                       )
-                                      (tee_local $1
+                                      (tee_local $0
                                         (i32.and
                                           (i32.shr_u
-                                            (tee_local $0
+                                            (tee_local $5
                                               (i32.shr_u
+                                                (get_local $0)
                                                 (get_local $1)
-                                                (get_local $4)
                                               )
                                             )
                                             (i32.const 1)
                                           )
-                                          (i32.const 2)
-                                        )
-                                      )
-                                    )
-                                    (tee_local $0
-                                      (i32.and
-                                        (i32.shr_u
-                                          (tee_local $5
-                                            (i32.shr_u
-                                              (get_local $0)
-                                              (get_local $1)
-                                            )
-                                          )
                                           (i32.const 1)
                                         )
-                                        (i32.const 1)
                                       )
                                     )
+                                    (i32.shr_u
+                                      (get_local $5)
+                                      (get_local $0)
+                                    )
                                   )
-                                  (i32.shr_u
-                                    (get_local $5)
-                                    (get_local $0)
-                                  )
+                                  (i32.const 2)
                                 )
-                                (i32.const 2)
+                                (i32.const 1512)
                               )
                             )
                           )
@@ -1463,106 +1466,109 @@
                 (block $label$break$a
                   (if
                     (tee_local $12
-                      (i32.load offset=1512
-                        (i32.shl
-                          (tee_local $27
-                            (if i32
-                              (tee_local $7
-                                (i32.shr_u
-                                  (get_local $1)
-                                  (i32.const 8)
-                                )
-                              )
+                      (i32.load
+                        (i32.add
+                          (i32.shl
+                            (tee_local $27
                               (if i32
-                                (i32.gt_u
-                                  (get_local $0)
-                                  (i32.const 16777215)
+                                (tee_local $7
+                                  (i32.shr_u
+                                    (get_local $1)
+                                    (i32.const 8)
+                                  )
                                 )
-                                (i32.const 31)
-                                (i32.or
-                                  (i32.and
-                                    (i32.shr_u
-                                      (get_local $0)
-                                      (i32.add
-                                        (tee_local $12
-                                          (i32.add
-                                            (i32.sub
-                                              (i32.const 14)
-                                              (i32.or
+                                (if i32
+                                  (i32.gt_u
+                                    (get_local $0)
+                                    (i32.const 16777215)
+                                  )
+                                  (i32.const 31)
+                                  (i32.or
+                                    (i32.and
+                                      (i32.shr_u
+                                        (get_local $0)
+                                        (i32.add
+                                          (tee_local $12
+                                            (i32.add
+                                              (i32.sub
+                                                (i32.const 14)
                                                 (i32.or
-                                                  (tee_local $7
-                                                    (i32.and
-                                                      (i32.shr_u
-                                                        (i32.add
-                                                          (tee_local $16
-                                                            (i32.shl
-                                                              (get_local $7)
-                                                              (tee_local $1
-                                                                (i32.and
-                                                                  (i32.shr_u
-                                                                    (i32.add
-                                                                      (get_local $7)
-                                                                      (i32.const 1048320)
+                                                  (i32.or
+                                                    (tee_local $7
+                                                      (i32.and
+                                                        (i32.shr_u
+                                                          (i32.add
+                                                            (tee_local $16
+                                                              (i32.shl
+                                                                (get_local $7)
+                                                                (tee_local $1
+                                                                  (i32.and
+                                                                    (i32.shr_u
+                                                                      (i32.add
+                                                                        (get_local $7)
+                                                                        (i32.const 1048320)
+                                                                      )
+                                                                      (i32.const 16)
                                                                     )
-                                                                    (i32.const 16)
+                                                                    (i32.const 8)
                                                                   )
-                                                                  (i32.const 8)
                                                                 )
                                                               )
                                                             )
+                                                            (i32.const 520192)
                                                           )
-                                                          (i32.const 520192)
+                                                          (i32.const 16)
                                                         )
-                                                        (i32.const 16)
+                                                        (i32.const 4)
                                                       )
-                                                      (i32.const 4)
                                                     )
+                                                    (get_local $1)
                                                   )
-                                                  (get_local $1)
-                                                )
-                                                (tee_local $16
-                                                  (i32.and
-                                                    (i32.shr_u
-                                                      (i32.add
-                                                        (tee_local $14
-                                                          (i32.shl
-                                                            (get_local $16)
-                                                            (get_local $7)
+                                                  (tee_local $16
+                                                    (i32.and
+                                                      (i32.shr_u
+                                                        (i32.add
+                                                          (tee_local $14
+                                                            (i32.shl
+                                                              (get_local $16)
+                                                              (get_local $7)
+                                                            )
                                                           )
+                                                          (i32.const 245760)
                                                         )
-                                                        (i32.const 245760)
+                                                        (i32.const 16)
                                                       )
-                                                      (i32.const 16)
+                                                      (i32.const 2)
                                                     )
-                                                    (i32.const 2)
                                                   )
                                                 )
                                               )
-                                            )
-                                            (i32.shr_u
-                                              (i32.shl
-                                                (get_local $14)
-                                                (get_local $16)
+                                              (i32.shr_u
+                                                (i32.shl
+                                                  (get_local $14)
+                                                  (get_local $16)
+                                                )
+                                                (i32.const 15)
                                               )
-                                              (i32.const 15)
                                             )
                                           )
+                                          (i32.const 7)
                                         )
-                                        (i32.const 7)
                                       )
+                                      (i32.const 1)
+                                    )
+                                    (i32.shl
+                                      (get_local $12)
+                                      (i32.const 1)
                                     )
-                                    (i32.const 1)
-                                  )
-                                  (i32.shl
-                                    (get_local $12)
-                                    (i32.const 1)
                                   )
                                 )
+                                (i32.const 0)
                               )
-                              (i32.const 0)
                             )
+                            (i32.const 2)
                           )
-                          (i32.const 2)
+                          (i32.const 1512)
                         )
                       )
                     )
@@ -1802,80 +1808,83 @@
                               (i32.const 16)
                             )
                           )
-                          (i32.load offset=1512
-                            (i32.shl
-                              (i32.add
-                                (i32.or
+                          (i32.load
+                            (i32.add
+                              (i32.shl
+                                (i32.add
                                   (i32.or
                                     (i32.or
                                       (i32.or
-                                        (tee_local $12
+                                        (i32.or
+                                          (tee_local $12
+                                            (i32.and
+                                              (i32.shr_u
+                                                (tee_local $3
+                                                  (i32.shr_u
+                                                    (get_local $12)
+                                                    (get_local $4)
+                                                  )
+                                                )
+                                                (i32.const 5)
+                                              )
+                                              (i32.const 8)
+                                            )
+                                          )
+                                          (get_local $4)
+                                        )
+                                        (tee_local $3
                                           (i32.and
                                             (i32.shr_u
-                                              (tee_local $3
+                                              (tee_local $5
                                                 (i32.shr_u
+                                                  (get_local $3)
                                                   (get_local $12)
-                                                  (get_local $4)
                                                 )
                                               )
-                                              (i32.const 5)
+                                              (i32.const 2)
                                             )
-                                            (i32.const 8)
+                                            (i32.const 4)
                                           )
                                         )
-                                        (get_local $4)
                                       )
-                                      (tee_local $3
+                                      (tee_local $5
                                         (i32.and
                                           (i32.shr_u
-                                            (tee_local $5
+                                            (tee_local $8
                                               (i32.shr_u
+                                                (get_local $5)
                                                 (get_local $3)
-                                                (get_local $12)
                                               )
                                             )
-                                            (i32.const 2)
+                                            (i32.const 1)
                                           )
-                                          (i32.const 4)
+                                          (i32.const 2)
                                         )
                                       )
                                     )
-                                    (tee_local $5
+                                    (tee_local $8
                                       (i32.and
                                         (i32.shr_u
-                                          (tee_local $8
+                                          (tee_local $1
                                             (i32.shr_u
+                                              (get_local $8)
                                               (get_local $5)
-                                              (get_local $3)
                                             )
                                           )
                                           (i32.const 1)
                                         )
-                                        (i32.const 2)
-                                      )
-                                    )
-                                  )
-                                  (tee_local $8
-                                    (i32.and
-                                      (i32.shr_u
-                                        (tee_local $1
-                                          (i32.shr_u
-                                            (get_local $8)
-                                            (get_local $5)
-                                          )
-                                        )
                                         (i32.const 1)
                                       )
-                                      (i32.const 1)
                                     )
                                   )
+                                  (i32.shr_u
+                                    (get_local $1)
+                                    (get_local $8)
+                                  )
                                 )
-                                (i32.shr_u
-                                  (get_local $1)
-                                  (get_local $8)
-                                )
+                                (i32.const 2)
                               )
-                              (i32.const 2)
+                              (i32.const 1512)
                             )
                           )
                         )
diff --git a/test/memorygrowth.fromasm.imprecise b/test/memorygrowth.fromasm.imprecise
index 3d244ccad..4301d1105 100644
--- a/test/memorygrowth.fromasm.imprecise
+++ b/test/memorygrowth.fromasm.imprecise
@@ -721,80 +721,83 @@
                       (i32.and
                         (i32.load offset=4
                           (tee_local $14
-                            (i32.load offset=1512
-                              (i32.shl
-                                (i32.add
-                                  (i32.or
+                            (i32.load
+                              (i32.add
+                                (i32.shl
+                                  (i32.add
                                     (i32.or
                                       (i32.or
                                         (i32.or
-                                          (tee_local $8
+                                          (i32.or
+                                            (tee_local $8
+                                              (i32.and
+                                                (i32.shr_u
+                                                  (tee_local $4
+                                                    (i32.shr_u
+                                                      (get_local $8)
+                                                      (get_local $12)
+                                                    )
+                                                  )
+                                                  (i32.const 5)
+                                                )
+                                                (i32.const 8)
+                                              )
+                                            )
+                                            (get_local $12)
+                                          )
+                                          (tee_local $4
                                             (i32.and
                                               (i32.shr_u
-                                                (tee_local $4
+                                                (tee_local $1
                                                   (i32.shr_u
+                                                    (get_local $4)
                                                     (get_local $8)
-                                                    (get_local $12)
                                                   )
                                                 )
-                                                (i32.const 5)
+                                                (i32.const 2)
                                               )
-                                              (i32.const 8)
+                                              (i32.const 4)
                                             )
                                           )
-                                          (get_local $12)
                                         )
-                                        (tee_local $4
+                                        (tee_local $1
                                           (i32.and
                                             (i32.shr_u
-                                              (tee_local $1
+                                              (tee_local $0
                                                 (i32.shr_u
+                                                  (get_local $1)
                                                   (get_local $4)
-                                                  (get_local $8)
                                                 )
                                               )
-                                              (i32.const 2)
+                                              (i32.const 1)
                                             )
-                                            (i32.const 4)
+                                            (i32.const 2)
                                           )
                                         )
                                       )
-                                      (tee_local $1
+                                      (tee_local $0
                                         (i32.and
                                           (i32.shr_u
-                                            (tee_local $0
+                                            (tee_local $5
                                               (i32.shr_u
+                                                (get_local $0)
                                                 (get_local $1)
-                                                (get_local $4)
                                               )
                                             )
                                             (i32.const 1)
                                           )
-                                          (i32.const 2)
-                                        )
-                                      )
-                                    )
-                                    (tee_local $0
-                                      (i32.and
-                                        (i32.shr_u
-                                          (tee_local $5
-                                            (i32.shr_u
-                                              (get_local $0)
-                                              (get_local $1)
-                                            )
-                                          )
                                           (i32.const 1)
                                         )
-                                        (i32.const 1)
                                       )
                                     )
+                                    (i32.shr_u
+                                      (get_local $5)
+                                      (get_local $0)
+                                    )
                                   )
-                                  (i32.shr_u
-                                    (get_local $5)
-                                    (get_local $0)
-                                  )
+                                  (i32.const 2)
                                 )
-                                (i32.const 2)
+                                (i32.const 1512)
                               )
                             )
                           )
@@ -1461,106 +1464,109 @@
                 (block $label$break$a
                   (if
                     (tee_local $12
-                      (i32.load offset=1512
-                        (i32.shl
-                          (tee_local $27
-                            (if i32
-                              (tee_local $7
-                                (i32.shr_u
-                                  (get_local $1)
-                                  (i32.const 8)
-                                )
-                              )
+                      (i32.load
+                        (i32.add
+                          (i32.shl
+                            (tee_local $27
                               (if i32
-                                (i32.gt_u
-                                  (get_local $0)
-                                  (i32.const 16777215)
+                                (tee_local $7
+                                  (i32.shr_u
+                                    (get_local $1)
+                                    (i32.const 8)
+                                  )
                                 )
-                                (i32.const 31)
-                                (i32.or
-                                  (i32.and
-                                    (i32.shr_u
-                                      (get_local $0)
-                                      (i32.add
-                                        (tee_local $12
-                                          (i32.add
-                                            (i32.sub
-                                              (i32.const 14)
-                                              (i32.or
+                                (if i32
+                                  (i32.gt_u
+                                    (get_local $0)
+                                    (i32.const 16777215)
+                                  )
+                                  (i32.const 31)
+                                  (i32.or
+                                    (i32.and
+                                      (i32.shr_u
+                                        (get_local $0)
+                                        (i32.add
+                                          (tee_local $12
+                                            (i32.add
+                                              (i32.sub
+                                                (i32.const 14)
                                                 (i32.or
-                                                  (tee_local $7
-                                                    (i32.and
-                                                      (i32.shr_u
-                                                        (i32.add
-                                                          (tee_local $16
-                                                            (i32.shl
-                                                              (get_local $7)
-                                                              (tee_local $1
-                                                                (i32.and
-                                                                  (i32.shr_u
-                                                                    (i32.add
-                                                                      (get_local $7)
-                                                                      (i32.const 1048320)
+                                                  (i32.or
+                                                    (tee_local $7
+                                                      (i32.and
+                                                        (i32.shr_u
+                                                          (i32.add
+                                                            (tee_local $16
+                                                              (i32.shl
+                                                                (get_local $7)
+                                                                (tee_local $1
+                                                                  (i32.and
+                                                                    (i32.shr_u
+                                                                      (i32.add
+                                                                        (get_local $7)
+                                                                        (i32.const 1048320)
+                                                                      )
+                                                                      (i32.const 16)
                                                                     )
-                                                                    (i32.const 16)
+                                                                    (i32.const 8)
                                                                   )
-                                                                  (i32.const 8)
                                                                 )
                                                               )
                                                             )
+                                                            (i32.const 520192)
                                                           )
-                                                          (i32.const 520192)
+                                                          (i32.const 16)
                                                         )
-                                                        (i32.const 16)
+                                                        (i32.const 4)
                                                       )
-                                                      (i32.const 4)
                                                     )
+                                                    (get_local $1)
                                                   )
-                                                  (get_local $1)
-                                                )
-                                                (tee_local $16
-                                                  (i32.and
-                                                    (i32.shr_u
-                                                      (i32.add
-                                                        (tee_local $14
-                                                          (i32.shl
-                                                            (get_local $16)
-                                                            (get_local $7)
+                                                  (tee_local $16
+                                                    (i32.and
+                                                      (i32.shr_u
+                                                        (i32.add
+                                                          (tee_local $14
+                                                            (i32.shl
+                                                              (get_local $16)
+                                                              (get_local $7)
+                                                            )
                                                           )
+                                                          (i32.const 245760)
                                                         )
-                                                        (i32.const 245760)
+                                                        (i32.const 16)
                                                       )
-                                                      (i32.const 16)
+                                                      (i32.const 2)
                                                     )
-                                                    (i32.const 2)
                                                   )
                                                 )
                                               )
-                                            )
-                                            (i32.shr_u
-                                              (i32.shl
-                                                (get_local $14)
-                                                (get_local $16)
+                                              (i32.shr_u
+                                                (i32.shl
+                                                  (get_local $14)
+                                                  (get_local $16)
+                                                )
+                                                (i32.const 15)
                                               )
-                                              (i32.const 15)
                                             )
                                           )
+                                          (i32.const 7)
                                         )
-                                        (i32.const 7)
                                       )
+                                      (i32.const 1)
+                                    )
+                                    (i32.shl
+                                      (get_local $12)
+                                      (i32.const 1)
                                     )
-                                    (i32.const 1)
-                                  )
-                                  (i32.shl
-                                    (get_local $12)
-                                    (i32.const 1)
                                   )
                                 )
+                                (i32.const 0)
                               )
-                              (i32.const 0)
                             )
+                            (i32.const 2)
                           )
-                          (i32.const 2)
+                          (i32.const 1512)
                         )
                       )
                     )
@@ -1800,80 +1806,83 @@
                               (i32.const 16)
                             )
                           )
-                          (i32.load offset=1512
-                            (i32.shl
-                              (i32.add
-                                (i32.or
+                          (i32.load
+                            (i32.add
+                              (i32.shl
+                                (i32.add
                                   (i32.or
                                     (i32.or
                                       (i32.or
-                                        (tee_local $12
+                                        (i32.or
+                                          (tee_local $12
+                                            (i32.and
+                                              (i32.shr_u
+                                                (tee_local $3
+                                                  (i32.shr_u
+                                                    (get_local $12)
+                                                    (get_local $4)
+                                                  )
+                                                )
+                                                (i32.const 5)
+                                              )
+                                              (i32.const 8)
+                                            )
+                                          )
+                                          (get_local $4)
+                                        )
+                                        (tee_local $3
                                           (i32.and
                                             (i32.shr_u
-                                              (tee_local $3
+                                              (tee_local $5
                                                 (i32.shr_u
+                                                  (get_local $3)
                                                   (get_local $12)
-                                                  (get_local $4)
                                                 )
                                               )
-                                              (i32.const 5)
+                                              (i32.const 2)
                                             )
-                                            (i32.const 8)
+                                            (i32.const 4)
                                           )
                                         )
-                                        (get_local $4)
                                       )
-                                      (tee_local $3
+                                      (tee_local $5
                                         (i32.and
                                           (i32.shr_u
-                                            (tee_local $5
+                                            (tee_local $8
                                               (i32.shr_u
+                                                (get_local $5)
                                                 (get_local $3)
-                                                (get_local $12)
                                               )
                                             )
-                                            (i32.const 2)
+                                            (i32.const 1)
                                           )
-                                          (i32.const 4)
+                                          (i32.const 2)
                                         )
                                       )
                                     )
-                                    (tee_local $5
+                                    (tee_local $8
                                       (i32.and
                                         (i32.shr_u
-                                          (tee_local $8
+                                          (tee_local $1
                                             (i32.shr_u
+                                              (get_local $8)
                                               (get_local $5)
-                                              (get_local $3)
                                             )
                                           )
                                           (i32.const 1)
                                         )
-                                        (i32.const 2)
-                                      )
-                                    )
-                                  )
-                                  (tee_local $8
-                                    (i32.and
-                                      (i32.shr_u
-                                        (tee_local $1
-                                          (i32.shr_u
-                                            (get_local $8)
-                                            (get_local $5)
-                                          )
-                                        )
                                         (i32.const 1)
                                       )
-                                      (i32.const 1)
                                     )
                                   )
+                                  (i32.shr_u
+                                    (get_local $1)
+                                    (get_local $8)
+                                  )
                                 )
-                                (i32.shr_u
-                                  (get_local $1)
-                                  (get_local $8)
-                                )
+                                (i32.const 2)
                               )
-                              (i32.const 2)
+                              (i32.const 1512)
                             )
                           )
                         )
diff --git a/test/passes/optimize-instructions.txt b/test/passes/optimize-instructions.txt
index a8cf68dbd..f959c3cd6 100644
--- a/test/passes/optimize-instructions.txt
+++ b/test/passes/optimize-instructions.txt
@@ -415,20 +415,32 @@
     )
   )
   (func $load-off-2 (type $3) (param $0 i32) (result i32)
-    (i32.store
-      (i32.const 6)
+    (i32.store offset=2
+      (i32.add
+        (i32.const 1)
+        (i32.const 3)
+      )
       (get_local $0)
     )
-    (i32.store
-      (i32.const 6)
+    (i32.store offset=2
+      (i32.add
+        (i32.const 3)
+        (i32.const 1)
+      )
       (get_local $0)
     )
-    (i32.store offset=7
-      (get_local $0)
+    (i32.store offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const 5)
+      )
       (get_local $0)
     )
-    (i32.store offset=9
-      (get_local $0)
+    (i32.store offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const 7)
+      )
       (get_local $0)
     )
     (i32.store offset=2
@@ -445,12 +457,18 @@
       )
       (get_local $0)
     )
-    (i32.store
-      (i32.const 4)
+    (i32.store offset=2
+      (i32.add
+        (i32.const -15)
+        (i32.const 17)
+      )
       (get_local $0)
     )
-    (i32.store
-      (i32.const 0)
+    (i32.store offset=2
+      (i32.add
+        (i32.const -21)
+        (i32.const 19)
+      )
       (get_local $0)
     )
     (i32.store
@@ -462,18 +480,27 @@
       (get_local $0)
     )
     (drop
-      (i32.load
-        (i32.const 8)
+      (i32.load offset=2
+        (i32.add
+          (i32.const 2)
+          (i32.const 4)
+        )
       )
     )
     (drop
-      (i32.load
-        (i32.const 8)
+      (i32.load offset=2
+        (i32.add
+          (i32.const 4)
+          (i32.const 2)
+        )
       )
     )
     (drop
-      (i32.load offset=8
-        (get_local $0)
+      (i32.load offset=2
+        (i32.add
+          (get_local $0)
+          (i32.const 6)
+        )
       )
     )
     (drop
@@ -481,8 +508,11 @@
         (i32.const 10)
       )
     )
-    (i32.load offset=12
-      (get_local $0)
+    (i32.load offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const 10)
+      )
     )
   )
 )
diff --git a/test/passes/post-emscripten.txt b/test/passes/post-emscripten.txt
index a2acbbb28..830d4276e 100644
--- a/test/passes/post-emscripten.txt
+++ b/test/passes/post-emscripten.txt
@@ -1,6 +1,8 @@
 (module
   (type $0 (func (param i32)))
+  (type $1 (func (param i32) (result i32)))
   (memory $0 256 256)
+  (export "load-off-2" (func $load-off-2))
   (func $b0 (type $0) (param $x i32)
     (drop
       (i32.load offset=1
@@ -39,4 +41,75 @@
       )
     )
   )
+  (func $load-off-2 (type $1) (param $0 i32) (result i32)
+    (i32.store
+      (i32.const 6)
+      (get_local $0)
+    )
+    (i32.store
+      (i32.const 6)
+      (get_local $0)
+    )
+    (i32.store offset=7
+      (get_local $0)
+      (get_local $0)
+    )
+    (i32.store offset=9
+      (get_local $0)
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const -11)
+        (get_local $0)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const -13)
+      )
+      (get_local $0)
+    )
+    (i32.store
+      (i32.const 4)
+      (get_local $0)
+    )
+    (i32.store
+      (i32.const 0)
+      (get_local $0)
+    )
+    (i32.store
+      (i32.const 25)
+      (get_local $0)
+    )
+    (i32.store
+      (i32.const -23)
+      (get_local $0)
+    )
+    (drop
+      (i32.load
+        (i32.const 8)
+      )
+    )
+    (drop
+      (i32.load
+        (i32.const 8)
+      )
+    )
+    (drop
+      (i32.load offset=8
+        (get_local $0)
+      )
+    )
+    (drop
+      (i32.load
+        (i32.const 10)
+      )
+    )
+    (i32.load offset=12
+      (get_local $0)
+    )
+  )
 )
diff --git a/test/passes/post-emscripten.wast b/test/passes/post-emscripten.wast
index b554fea53..f262975f5 100644
--- a/test/passes/post-emscripten.wast
+++ b/test/passes/post-emscripten.wast
@@ -51,4 +51,105 @@
       )
     )
   )
+  (func $load-off-2 "load-off-2" (param $0 i32) (result i32)
+    (i32.store offset=2
+      (i32.add
+        (i32.const 1)
+        (i32.const 3)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const 3)
+        (i32.const 1)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const 5)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const 7)
+        (get_local $0)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const -11) ;; do not fold this!
+        (get_local $0)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (get_local $0)
+        (i32.const -13) ;; do not fold this!
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const -15)
+        (i32.const 17)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.add
+        (i32.const -21)
+        (i32.const 19)
+      )
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.const 23)
+      (get_local $0)
+    )
+    (i32.store offset=2
+      (i32.const -25)
+      (get_local $0)
+    )
+    (drop
+      (i32.load offset=2
+        (i32.add
+          (i32.const 2)
+          (i32.const 4)
+        )
+      )
+    )
+    (drop
+      (i32.load offset=2
+        (i32.add
+          (i32.const 4)
+          (i32.const 2)
+        )
+      )
+    )
+    (drop
+      (i32.load offset=2
+        (i32.add
+          (get_local $0)
+          (i32.const 6)
+        )
+      )
+    )
+    (drop
+      (i32.load offset=2
+        (i32.const 8)
+      )
+    )
+    (i32.load offset=2
+      (i32.add
+        (i32.const 10)
+        (get_local $0)
+      )
+    )
+  )
 )