fix stack popping in s2wasm

21 files changed, 242 insertions, 210 deletions

diff --git a/src/s2wasm.h b/src/s2wasm.h
index 759dcf44a..9a0ab69c5 100644
--- a/src/s2wasm.h
+++ b/src/s2wasm.h
@@ -111,6 +111,12 @@ private:
     return cashew::IString(str.c_str(), false);
   }
 
+  void skipToSep() {
+    while (*s && !isspace(*s) && *s != ',' && *s != ')') {
+      s++;
+    }
+  }
+
   Name getStrToSep() {
     std::string str;
     while (*s && !isspace(*s) && *s != ',' && *s != ')') {
@@ -345,17 +351,42 @@ private:
       //std::cerr << "pop " << ret << '\n';
       return ret;
     };
-    auto getInput = [&]() {
-      //dump("getinput");
-      if (match("$pop")) {
-        while (isdigit(*s)) s++;
-        return pop();
-      } else {
-        auto curr = allocator.alloc<GetLocal>();
-        curr->name = getStrToSep();
-        curr->type = localTypes[curr->name];
-        return (Expression*)curr;
+    auto getNumInputs = [&]() {
+      int ret = 1;
+      char *t = s;
+      while (*t != '\n') {
+        if (*t == ',') ret++;
+        t++;
       }
+      return ret;
+    };
+    auto getInputs = [&](int num) {
+      // we may have       $pop, $0, $pop, $1     etc., which are getlocals
+      // interleaved with stack pops, and the stack pops must be done in
+      // *reverse* order, i.e., that input should turn into
+      //         lastpop, getlocal(0), firstpop, getlocal(1)
+      std::vector<Expression*> inputs; // TODO: optimize (if .s format doesn't change)
+      inputs.resize(num);
+      for (int i = 0; i < num; i++) {
+        if (match("$pop")) {
+          skipToSep();
+          inputs[i] = nullptr;
+        } else {
+          auto curr = allocator.alloc<GetLocal>();
+          curr->name = getStrToSep();
+          curr->type = localTypes[curr->name];
+          inputs[i] = curr;
+        }
+        if (*s == ')') s++; // tolerate 0(argument) syntax, where we started at the 'a'
+        if (i < num - 1) skipComma();
+      }
+      for (int i = num-1; i >= 0; i--) {
+        if (inputs[i] == nullptr) inputs[i] = pop();
+      }
+      return inputs;
+    };
+    auto getInput = [&]() {
+      return getInputs(1)[0];
     };
     auto setOutput = [&](Expression* curr, Name assign) {
       if (assign.isNull() || assign.str[1] == 'd') { // discard
@@ -375,9 +406,9 @@ private:
       skipComma();
       auto curr = allocator.alloc<Binary>();
       curr->op = op;
-      curr->right = getInput();
-      skipComma();
-      curr->left = getInput();
+      auto inputs = getInputs(2);
+      curr->left = inputs[0];
+      curr->right = inputs[1];
       curr->finalize();
       assert(curr->type == type);
       setOutput(curr, assign);
@@ -417,7 +448,6 @@ private:
       curr->align = curr->bytes; // XXX
       mustMatch("(");
       curr->ptr = getInput();
-      mustMatch(")");
       setOutput(curr, assign);
     };
     auto makeStore = [&](WasmType type) {
@@ -426,26 +456,23 @@ private:
       curr->type = type;
       int32_t bytes = getInt();
       curr->bytes = bytes > 0 ? bytes : getWasmTypeSize(type);
+      curr->align = curr->bytes; // XXX
       Name assign = getAssign();
       curr->offset = getInt();
-      curr->align = curr->bytes; // XXX
       mustMatch("(");
-      curr->ptr = getInput();
-      mustMatch(")");
-      skipComma();
-      curr->value = getInput();
+      auto inputs = getInputs(2);
+      curr->ptr = inputs[0];
+      curr->value = inputs[1];
       setOutput(curr, assign);
     };
     auto makeSelect = [&](WasmType type) {
       Name assign = getAssign();
       skipComma();
       auto curr = allocator.alloc<Select>();
-      curr->condition = getInput();
-      skipComma();
-      curr->ifTrue = getInput();
-      skipComma();
-      curr->ifFalse = getInput();
-      skipComma();
+      auto inputs = getInputs(3);
+      curr->condition = inputs[0];
+      curr->ifTrue = inputs[1];
+      curr->ifFalse = inputs[2];
       curr->type = type;
       setOutput(curr, assign);
     };
@@ -477,9 +504,14 @@ private:
         }
       }
       curr->type = type;
-      while (1) {
-        if (!skipComma()) break;
-        curr->operands.push_back(getInput());
+      skipWhitespace();
+      if (*s == ',') {
+        skipComma();
+        int num = getNumInputs();
+        auto inputs = getInputs(num);
+        for (int i = 0; i < num; i++) {
+          curr->operands.push_back(inputs[i]);
+        }
       }
       std::reverse(curr->operands.begin(), curr->operands.end());
       setOutput(curr, assign);
diff --git a/test/dot_s/basics.wast b/test/dot_s/basics.wast
index 1a6af338c..7c7e43ec3 100644
--- a/test/dot_s/basics.wast
+++ b/test/dot_s/basics.wast
@@ -13,20 +13,20 @@
             (br_if
               (i32.ne
                 (i32.sub
+                  (get_local $$0)
                   (i32.and
                     (i32.add
+                      (get_local $$0)
                       (i32.shr_u
                         (i32.shr_s
-                          (i32.const 31)
                           (get_local $$0)
+                          (i32.const 31)
                         )
                         (i32.const 30)
                       )
-                      (get_local $$0)
                     )
                     (i32.const -4)
                   )
-                  (get_local $$0)
                 )
                 (i32.const 1)
               )
@@ -36,19 +36,19 @@
               (block
                 (set_local $$0
                   (i32.add
-                    (get_local $$0)
                     (i32.gt_s
-                      (i32.const 10)
                       (get_local $$0)
+                      (i32.const 10)
                     )
+                    (get_local $$0)
                   )
                 )
                 (block $BB0_3
                   (br_if
                     (i32.ne
                       (i32.rem_s
-                        (i32.const 5)
                         (get_local $$0)
+                        (i32.const 5)
                       )
                       (i32.const 3)
                     )
@@ -56,19 +56,19 @@
                   )
                   (set_local $$0
                     (i32.add
-                      (get_local $$0)
                       (i32.rem_s
-                        (i32.const 111)
                         (get_local $$0)
+                        (i32.const 111)
                       )
+                      (get_local $$0)
                     )
                   )
                 )
                 (br_if
                   (i32.eq
                     (i32.rem_s
-                      (i32.const 7)
                       (get_local $$0)
+                      (i32.const 7)
                     )
                     (i32.const 0)
                   )
@@ -80,8 +80,8 @@
           )
           (set_local $$0
             (i32.add
-              (i32.const -12)
               (get_local $$0)
+              (i32.const -12)
             )
           )
         )
diff --git a/test/dot_s/cfg-stackify.wast b/test/dot_s/cfg-stackify.wast
index 84440a122..2f6034844 100644
--- a/test/dot_s/cfg-stackify.wast
+++ b/test/dot_s/cfg-stackify.wast
@@ -36,14 +36,14 @@
           (block
             (set_local $$1
               (i32.add
-                (i32.const 1)
                 (get_local $$1)
+                (i32.const 1)
               )
             )
             (br_if
               (i32.ge_s
-                (get_local $$0)
                 (get_local $$1)
+                (get_local $$0)
               )
               $BB0_3
             )
@@ -66,14 +66,14 @@
           (block
             (set_local $$1
               (i32.add
-                (i32.const 1)
                 (get_local $$1)
+                (i32.const 1)
               )
             )
             (br_if
               (i32.ge_s
-                (get_local $$0)
                 (get_local $$1)
+                (get_local $$0)
               )
               $BB1_3
             )
@@ -91,8 +91,8 @@
         (block $BB2_2
           (br_if
             (i32.lt_s
-              (i32.const 1)
               (get_local $$1)
+              (i32.const 1)
             )
             $BB2_2
           )
@@ -109,14 +109,14 @@
               )
               (set_local $$1
                 (i32.add
-                  (i32.const -1)
                   (get_local $$1)
+                  (i32.const -1)
                 )
               )
               (set_local $$0
                 (i32.add
-                  (i32.const 8)
                   (get_local $$0)
+                  (i32.const 8)
                 )
               )
               (br_if
@@ -296,8 +296,8 @@
             )
             (br_if
               (i32.eq
-                (i32.const 0)
                 (get_local $$1)
+                (i32.const 0)
               )
               $BB8_1
             )
@@ -494,23 +494,23 @@
             (block $BB13_4
               (br_if
                 (i32.gt_s
-                  (i32.const 3)
                   (get_local $$0)
+                  (i32.const 3)
                 )
                 $BB13_4
               )
               (block $BB13_3
                 (br_if
                   (i32.eq
-                    (i32.const 0)
                     (get_local $$0)
+                    (i32.const 0)
                   )
                   $BB13_3
                 )
                 (br_if
                   (i32.ne
-                    (i32.const 2)
                     (get_local $$0)
+                    (i32.const 2)
                   )
                   $BB13_7
                 )
@@ -519,15 +519,15 @@
             )
             (br_if
               (i32.eq
-                (i32.const 4)
                 (get_local $$0)
+                (i32.const 4)
               )
               $BB13_8
             )
             (br_if
               (i32.ne
-                (i32.const 622)
                 (get_local $$0)
+                (i32.const 622)
               )
               $BB13_7
             )
@@ -562,8 +562,8 @@
               (br_if
                 (i32.eq
                   (i32.and
-                    (get_local $$2)
                     (get_local $$0)
+                    (get_local $$2)
                   )
                   (i32.const 0)
                 )
@@ -571,11 +571,11 @@
               )
               (br_if
                 (i32.and
+                  (get_local $$1)
                   (i32.store align=4
                     (get_local $$3)
                     (get_local $$2)
                   )
-                  (get_local $$1)
                 )
                 $BB14_1
               )
@@ -618,8 +618,8 @@
                 (br_if
                   (i32.eq
                     (i32.and
-                      (get_local $$3)
                       (get_local $$0)
+                      (get_local $$3)
                     )
                     (i32.const 0)
                   )
@@ -631,14 +631,14 @@
                 )
                 (set_local $$4
                   (i32.and
-                    (get_local $$3)
                     (get_local $$1)
+                    (get_local $$3)
                   )
                 )
                 (br_if
                   (i32.eq
-                    (i32.const 0)
                     (get_local $$4)
+                    (i32.const 0)
                   )
                   $BB15_5
                 )
@@ -695,8 +695,8 @@
           (block $BB16_4
             (br_if
               (i32.and
-                (get_local $$3)
                 (get_local $$0)
+                (get_local $$3)
               )
               $BB16_4
             )
@@ -706,8 +706,8 @@
             )
             (br_if
               (i32.and
-                (get_local $$3)
                 (get_local $$1)
+                (get_local $$3)
               )
               $BB16_1
             )
@@ -723,8 +723,8 @@
           )
           (br_if
             (i32.and
-              (get_local $$3)
               (get_local $$1)
+              (get_local $$3)
             )
             $BB16_1
           )
@@ -748,15 +748,15 @@
           (block $BB17_3
             (br_if
               (i32.eq
-                (i32.const 0)
                 (get_local $$0)
+                (i32.const 0)
               )
               $BB17_3
             )
             (br_if
               (i32.eq
-                (i32.const 0)
                 (get_local $$0)
+                (i32.const 0)
               )
               $BB17_1
             )
@@ -799,8 +799,8 @@
             (br_if
               (i32.eq
                 (i32.and
-                  (get_local $$1)
                   (call_import $a)
+                  (get_local $$1)
                 )
                 (i32.const 0)
               )
@@ -816,8 +816,8 @@
                   (br_if
                     (i32.eq
                       (i32.and
-                        (get_local $$1)
                         (call_import $a)
+                        (get_local $$1)
                       )
                       (i32.const 0)
                     )
@@ -829,8 +829,8 @@
                   )
                   (br_if
                     (i32.and
-                      (get_local $$1)
                       (call_import $a)
+                      (get_local $$1)
                     )
                     $BB18_2
                   )
@@ -842,8 +842,8 @@
                 )
                 (br_if
                   (i32.and
-                    (get_local $$1)
                     (call_import $a)
+                    (get_local $$1)
                   )
                   $BB18_2
                 )
@@ -908,8 +908,8 @@
                       )
                       (br_if
                         (i32.gt_u
-                          (i32.const 4)
                           (get_local $$2)
+                          (i32.const 4)
                         )
                         $BB19_1
                       )
@@ -992,8 +992,8 @@
               )
               (br_if
                 (i32.eq
-                  (i32.const 0)
                   (get_local $$0)
+                  (i32.const 0)
                 )
                 $BB20_7
               )
@@ -1034,22 +1034,22 @@
                 (block $BB21_4
                   (br_if
                     (i32.gt_s
-                      (i32.const 103)
                       (get_local $$1)
+                      (i32.const 103)
                     )
                     $BB21_4
                   )
                   (br_if
                     (i32.eq
-                      (i32.const 42)
                       (get_local $$1)
+                      (i32.const 42)
                     )
                     $BB21_7
                   )
                   (br_if
                     (i32.eq
-                      (i32.const 76)
                       (get_local $$1)
+                      (i32.const 76)
                     )
                     $BB21_7
                   )
@@ -1057,15 +1057,15 @@
                 )
                 (br_if
                   (i32.eq
-                    (i32.const 108)
                     (get_local $$1)
+                    (i32.const 108)
                   )
                   $BB21_7
                 )
                 (br_if
                   (i32.eq
-                    (i32.const 104)
                     (get_local $$1)
+                    (i32.const 104)
                   )
                   $BB21_7
                 )
@@ -1074,8 +1074,8 @@
             )
             (set_local $$0
               (i32.add
-                (i32.const 1)
                 (get_local $$0)
+                (i32.const 1)
               )
             )
             (br $BB21_1)
diff --git a/test/dot_s/comparisons_f32.wast b/test/dot_s/comparisons_f32.wast
index 75e3af9e5..5b00f7e60 100644
--- a/test/dot_s/comparisons_f32.wast
+++ b/test/dot_s/comparisons_f32.wast
@@ -55,8 +55,8 @@
       (block
         (br $fake_return_waka123
           (f32.eq
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -67,8 +67,8 @@
       (block
         (br $fake_return_waka123
           (f32.ne
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -79,8 +79,8 @@
       (block
         (br $fake_return_waka123
           (f32.lt
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -91,8 +91,8 @@
       (block
         (br $fake_return_waka123
           (f32.le
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -103,8 +103,8 @@
       (block
         (br $fake_return_waka123
           (f32.gt
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -115,8 +115,8 @@
       (block
         (br $fake_return_waka123
           (f32.ge
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -128,8 +128,8 @@
         (br $fake_return_waka123
           (i32.or
             (f32.eq
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f32.ne
@@ -152,8 +152,8 @@
         (br $fake_return_waka123
           (i32.and
             (f32.ne
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.and
               (f32.eq
@@ -176,8 +176,8 @@
         (br $fake_return_waka123
           (i32.or
             (f32.lt
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f32.ne
@@ -200,8 +200,8 @@
         (br $fake_return_waka123
           (i32.or
             (f32.le
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f32.ne
@@ -224,8 +224,8 @@
         (br $fake_return_waka123
           (i32.or
             (f32.gt
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f32.ne
@@ -248,8 +248,8 @@
         (br $fake_return_waka123
           (i32.or
             (f32.ge
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f32.ne
diff --git a/test/dot_s/comparisons_f64.wast b/test/dot_s/comparisons_f64.wast
index 487d1f298..8f01136b0 100644
--- a/test/dot_s/comparisons_f64.wast
+++ b/test/dot_s/comparisons_f64.wast
@@ -55,8 +55,8 @@
       (block
         (br $fake_return_waka123
           (f64.eq
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -67,8 +67,8 @@
       (block
         (br $fake_return_waka123
           (f64.ne
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -79,8 +79,8 @@
       (block
         (br $fake_return_waka123
           (f64.lt
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -91,8 +91,8 @@
       (block
         (br $fake_return_waka123
           (f64.le
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -103,8 +103,8 @@
       (block
         (br $fake_return_waka123
           (f64.gt
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -115,8 +115,8 @@
       (block
         (br $fake_return_waka123
           (f64.ge
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -128,8 +128,8 @@
         (br $fake_return_waka123
           (i32.or
             (f64.eq
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f64.ne
@@ -152,8 +152,8 @@
         (br $fake_return_waka123
           (i32.and
             (f64.ne
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.and
               (f64.eq
@@ -176,8 +176,8 @@
         (br $fake_return_waka123
           (i32.or
             (f64.lt
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f64.ne
@@ -200,8 +200,8 @@
         (br $fake_return_waka123
           (i32.or
             (f64.le
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f64.ne
@@ -224,8 +224,8 @@
         (br $fake_return_waka123
           (i32.or
             (f64.gt
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f64.ne
@@ -248,8 +248,8 @@
         (br $fake_return_waka123
           (i32.or
             (f64.ge
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
             (i32.or
               (f64.ne
diff --git a/test/dot_s/comparisons_i32.wast b/test/dot_s/comparisons_i32.wast
index 84eed3086..f759d5a6a 100644
--- a/test/dot_s/comparisons_i32.wast
+++ b/test/dot_s/comparisons_i32.wast
@@ -15,8 +15,8 @@
       (block
         (br $fake_return_waka123
           (i32.eq
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -27,8 +27,8 @@
       (block
         (br $fake_return_waka123
           (i32.ne
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -39,8 +39,8 @@
       (block
         (br $fake_return_waka123
           (i32.lt_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -51,8 +51,8 @@
       (block
         (br $fake_return_waka123
           (i32.le_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -63,8 +63,8 @@
       (block
         (br $fake_return_waka123
           (i32.lt_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -75,8 +75,8 @@
       (block
         (br $fake_return_waka123
           (i32.le_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -87,8 +87,8 @@
       (block
         (br $fake_return_waka123
           (i32.gt_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -99,8 +99,8 @@
       (block
         (br $fake_return_waka123
           (i32.ge_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -111,8 +111,8 @@
       (block
         (br $fake_return_waka123
           (i32.gt_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -123,8 +123,8 @@
       (block
         (br $fake_return_waka123
           (i32.ge_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/comparisons_i64.wast b/test/dot_s/comparisons_i64.wast
index c14dca3c8..35479ecc2 100644
--- a/test/dot_s/comparisons_i64.wast
+++ b/test/dot_s/comparisons_i64.wast
@@ -15,8 +15,8 @@
       (block
         (br $fake_return_waka123
           (i64.eq
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -27,8 +27,8 @@
       (block
         (br $fake_return_waka123
           (i64.ne
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -39,8 +39,8 @@
       (block
         (br $fake_return_waka123
           (i64.lt_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -51,8 +51,8 @@
       (block
         (br $fake_return_waka123
           (i64.le_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -63,8 +63,8 @@
       (block
         (br $fake_return_waka123
           (i64.lt_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -75,8 +75,8 @@
       (block
         (br $fake_return_waka123
           (i64.le_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -87,8 +87,8 @@
       (block
         (br $fake_return_waka123
           (i64.gt_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -99,8 +99,8 @@
       (block
         (br $fake_return_waka123
           (i64.ge_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -111,8 +111,8 @@
       (block
         (br $fake_return_waka123
           (i64.gt_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -123,8 +123,8 @@
       (block
         (br $fake_return_waka123
           (i64.ge_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/dead-vreg.wast b/test/dot_s/dead-vreg.wast
index b5559e1ed..3445f5800 100644
--- a/test/dot_s/dead-vreg.wast
+++ b/test/dot_s/dead-vreg.wast
@@ -17,15 +17,15 @@
         (block $BB0_5
           (br_if
             (i32.lt_s
-              (get_local $$4)
               (get_local $$2)
+              (get_local $$4)
             )
             $BB0_5
           )
           (set_local $$3
             (i32.shl
-              (i32.const 2)
               (get_local $$1)
+              (i32.const 2)
             )
           )
           (set_local $$5
@@ -48,8 +48,8 @@
               (block $BB0_4
                 (br_if
                   (i32.lt_s
-                    (get_local $$4)
                     (get_local $$1)
+                    (get_local $$4)
                   )
                   $BB0_4
                 )
@@ -57,8 +57,8 @@
                   (block
                     (set_local $$9
                       (i32.add
-                        (i32.const -1)
                         (get_local $$9)
+                        (i32.const -1)
                       )
                     )
                     (i32.store align=4
@@ -67,14 +67,14 @@
                     )
                     (set_local $$8
                       (i32.add
-                        (i32.const 4)
                         (get_local $$8)
+                        (i32.const 4)
                       )
                     )
                     (set_local $$7
                       (i32.add
-                        (get_local $$6)
                         (get_local $$7)
+                        (get_local $$6)
                       )
                     )
                     (br_if
@@ -86,20 +86,20 @@
               )
               (set_local $$6
                 (i32.add
-                  (get_local $$4)
                   (get_local $$6)
+                  (get_local $$4)
                 )
               )
               (set_local $$0
                 (i32.add
-                  (get_local $$3)
                   (get_local $$0)
+                  (get_local $$3)
                 )
               )
               (br_if
                 (i32.ne
-                  (get_local $$2)
                   (get_local $$6)
+                  (get_local $$2)
                 )
                 $BB0_2
               )
diff --git a/test/dot_s/f32.wast b/test/dot_s/f32.wast
index 21141a396..fa4f87de8 100644
--- a/test/dot_s/f32.wast
+++ b/test/dot_s/f32.wast
@@ -22,8 +22,8 @@
       (block
         (br $fake_return_waka123
           (f32.add
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -34,8 +34,8 @@
       (block
         (br $fake_return_waka123
           (f32.sub
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -46,8 +46,8 @@
       (block
         (br $fake_return_waka123
           (f32.mul
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -58,8 +58,8 @@
       (block
         (br $fake_return_waka123
           (f32.div
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -92,8 +92,8 @@
       (block
         (br $fake_return_waka123
           (f32.copysign
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -170,8 +170,8 @@
       (block
         (br $fake_return_waka123
           (f32.min
-            (f32.const 0)
             (get_local $$0)
+            (f32.const 0)
           )
         )
       )
@@ -182,8 +182,8 @@
       (block
         (br $fake_return_waka123
           (f32.max
-            (f32.const 0)
             (get_local $$0)
+            (f32.const 0)
           )
         )
       )
diff --git a/test/dot_s/f64.wast b/test/dot_s/f64.wast
index 20866426f..afb3ca9df 100644
--- a/test/dot_s/f64.wast
+++ b/test/dot_s/f64.wast
@@ -22,8 +22,8 @@
       (block
         (br $fake_return_waka123
           (f64.add
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -34,8 +34,8 @@
       (block
         (br $fake_return_waka123
           (f64.sub
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -46,8 +46,8 @@
       (block
         (br $fake_return_waka123
           (f64.mul
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -58,8 +58,8 @@
       (block
         (br $fake_return_waka123
           (f64.div
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -92,8 +92,8 @@
       (block
         (br $fake_return_waka123
           (f64.copysign
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -170,8 +170,8 @@
       (block
         (br $fake_return_waka123
           (f64.min
-            (f64.const 0)
             (get_local $$0)
+            (f64.const 0)
           )
         )
       )
@@ -182,8 +182,8 @@
       (block
         (br $fake_return_waka123
           (f64.max
-            (f64.const 0)
             (get_local $$0)
+            (f64.const 0)
           )
         )
       )
diff --git a/test/dot_s/func.wast b/test/dot_s/func.wast
index d0f20d61c..ffb938022 100644
--- a/test/dot_s/func.wast
+++ b/test/dot_s/func.wast
@@ -49,8 +49,8 @@
           (br_if
             (i32.eq
               (i32.and
-                (get_local $$1)
                 (get_local $$0)
+                (get_local $$1)
               )
               (i32.const 0)
             )
diff --git a/test/dot_s/i32.wast b/test/dot_s/i32.wast
index 827ba8725..a1a690730 100644
--- a/test/dot_s/i32.wast
+++ b/test/dot_s/i32.wast
@@ -23,8 +23,8 @@
       (block
         (br $fake_return_waka123
           (i32.add
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -35,8 +35,8 @@
       (block
         (br $fake_return_waka123
           (i32.sub
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -47,8 +47,8 @@
       (block
         (br $fake_return_waka123
           (i32.mul
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -59,8 +59,8 @@
       (block
         (br $fake_return_waka123
           (i32.div_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -71,8 +71,8 @@
       (block
         (br $fake_return_waka123
           (i32.div_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -83,8 +83,8 @@
       (block
         (br $fake_return_waka123
           (i32.rem_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -95,8 +95,8 @@
       (block
         (br $fake_return_waka123
           (i32.rem_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -107,8 +107,8 @@
       (block
         (br $fake_return_waka123
           (i32.and
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -119,8 +119,8 @@
       (block
         (br $fake_return_waka123
           (i32.or
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -131,8 +131,8 @@
       (block
         (br $fake_return_waka123
           (i32.xor
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -143,8 +143,8 @@
       (block
         (br $fake_return_waka123
           (i32.shl
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -155,8 +155,8 @@
       (block
         (br $fake_return_waka123
           (i32.shr_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -167,8 +167,8 @@
       (block
         (br $fake_return_waka123
           (i32.shr_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/i64.wast b/test/dot_s/i64.wast
index a3befb718..cc52a53b6 100644
--- a/test/dot_s/i64.wast
+++ b/test/dot_s/i64.wast
@@ -23,8 +23,8 @@
       (block
         (br $fake_return_waka123
           (i64.add
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -35,8 +35,8 @@
       (block
         (br $fake_return_waka123
           (i64.sub
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -47,8 +47,8 @@
       (block
         (br $fake_return_waka123
           (i64.mul
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -59,8 +59,8 @@
       (block
         (br $fake_return_waka123
           (i64.div_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -71,8 +71,8 @@
       (block
         (br $fake_return_waka123
           (i64.div_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -83,8 +83,8 @@
       (block
         (br $fake_return_waka123
           (i64.rem_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -95,8 +95,8 @@
       (block
         (br $fake_return_waka123
           (i64.rem_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -107,8 +107,8 @@
       (block
         (br $fake_return_waka123
           (i64.and
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -119,8 +119,8 @@
       (block
         (br $fake_return_waka123
           (i64.or
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -131,8 +131,8 @@
       (block
         (br $fake_return_waka123
           (i64.xor
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -143,8 +143,8 @@
       (block
         (br $fake_return_waka123
           (i64.shl
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -155,8 +155,8 @@
       (block
         (br $fake_return_waka123
           (i64.shr_u
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
@@ -167,8 +167,8 @@
       (block
         (br $fake_return_waka123
           (i64.shr_s
-            (get_local $$1)
             (get_local $$0)
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/legalize.wast b/test/dot_s/legalize.wast
index 97ba128ab..761d87ed4 100644
--- a/test/dot_s/legalize.wast
+++ b/test/dot_s/legalize.wast
@@ -10,11 +10,11 @@
       (block
         (br $fake_return_waka123
           (i32.shl
+            (get_local $$0)
             (i32.and
-              (i32.const 7)
               (get_local $$1)
+              (i32.const 7)
             )
-            (get_local $$0)
           )
         )
       )
@@ -25,11 +25,11 @@
       (block
         (br $fake_return_waka123
           (i64.shl
+            (get_local $$0)
             (i64.and
-              (i64.const 9007199254740991)
               (get_local $$1)
+              (i64.const 9007199254740991)
             )
-            (get_local $$0)
           )
         )
       )
@@ -44,11 +44,11 @@
         )
         (br $fake_return_waka123
           (i64.shr_s
-            (get_local $$1)
             (i64.shl
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/load-store-i1.wast b/test/dot_s/load-store-i1.wast
index c8728f0ce..7acb8faea 100644
--- a/test/dot_s/load-store-i1.wast
+++ b/test/dot_s/load-store-i1.wast
@@ -26,13 +26,13 @@
         )
         (br $fake_return_waka123
           (i32.shr_s
-            (get_local $$1)
             (i32.shl
-              (get_local $$1)
               (i32.load align=8
                 (get_local $$0)
               )
+              (get_local $$1)
             )
+            (get_local $$1)
           )
         )
       )
@@ -58,13 +58,13 @@
         )
         (br $fake_return_waka123
           (i64.shr_s
-            (get_local $$1)
             (i64.shl
-              (get_local $$1)
               (i64.load align=8
                 (get_local $$0)
               )
+              (get_local $$1)
             )
+            (get_local $$1)
           )
         )
       )
@@ -76,8 +76,8 @@
         (i32.store align=8
           (get_local $$0)
           (i32.and
-            (i32.const 1)
             (get_local $$1)
+            (i32.const 1)
           )
         )
         (br $fake_return_waka123)
@@ -90,8 +90,8 @@
         (i64.store align=8
           (get_local $$0)
           (i64.and
-            (i64.const 1)
             (get_local $$1)
+            (i64.const 1)
           )
         )
         (br $fake_return_waka123)
diff --git a/test/dot_s/phi.wast b/test/dot_s/phi.wast
index 896aad133..2fb0ddf1e 100644
--- a/test/dot_s/phi.wast
+++ b/test/dot_s/phi.wast
@@ -8,15 +8,15 @@
         (block $BB0_2
           (br_if
             (i32.gt_s
-              (i32.const -1)
               (get_local $$0)
+              (i32.const -1)
             )
             $BB0_2
           )
           (set_local $$0
             (i32.div_s
-              (i32.const 3)
               (get_local $$0)
+              (i32.const 3)
             )
           )
         )
@@ -50,8 +50,8 @@
           (block
             (set_local $$5
               (i32.add
-                (get_local $$2)
                 (get_local $$5)
+                (get_local $$2)
               )
             )
             (set_local $$1
@@ -65,8 +65,8 @@
             )
             (br_if
               (i32.lt_s
-                (get_local $$0)
                 (get_local $$5)
+                (get_local $$0)
               )
               $BB1_1
             )
diff --git a/test/dot_s/reg-stackify.wast b/test/dot_s/reg-stackify.wast
index cb9b720d9..74d3d94b5 100644
--- a/test/dot_s/reg-stackify.wast
+++ b/test/dot_s/reg-stackify.wast
@@ -81,29 +81,29 @@
         (block $BB4_2
           (br_if
             (i32.ne
-              (get_local $$4)
               (i32.xor
                 (i32.xor
                   (i32.lt_s
-                    (get_local $$4)
                     (get_local $$0)
+                    (get_local $$4)
                   )
                   (i32.lt_s
-                    (get_local $$5)
                     (get_local $$1)
+                    (get_local $$5)
                   )
                 )
                 (i32.xor
                   (i32.lt_s
-                    (get_local $$4)
                     (get_local $$2)
+                    (get_local $$4)
                   )
                   (i32.lt_s
-                    (get_local $$5)
                     (get_local $$3)
+                    (get_local $$5)
                   )
                 )
               )
+              (get_local $$4)
             )
             $BB4_2
           )
diff --git a/test/dot_s/signext-zeroext.wast b/test/dot_s/signext-zeroext.wast
index 4db88e8ea..ddd3e85af 100644
--- a/test/dot_s/signext-zeroext.wast
+++ b/test/dot_s/signext-zeroext.wast
@@ -13,11 +13,11 @@
         )
         (br $fake_return_waka123
           (i32.shr_s
-            (get_local $$1)
             (i32.shl
-              (get_local $$1)
               (get_local $$0)
+              (get_local $$1)
             )
+            (get_local $$1)
           )
         )
       )
@@ -28,8 +28,8 @@
       (block
         (br $fake_return_waka123
           (i32.and
-            (i32.const 255)
             (get_local $$0)
+            (i32.const 255)
           )
         )
       )
@@ -41,8 +41,8 @@
         (br $fake_return_waka123
           (call $z2s_func
             (i32.and
-              (i32.const 255)
               (get_local $$0)
+              (i32.const 255)
             )
           )
         )
@@ -58,19 +58,19 @@
         )
         (br $fake_return_waka123
           (i32.shr_s
-            (get_local $$1)
             (i32.shl
-              (get_local $$1)
               (call $s2z_func
                 (i32.shr_s
-                  (get_local $$1)
                   (i32.shl
-                    (get_local $$1)
                     (get_local $$0)
+                    (get_local $$1)
                   )
+                  (get_local $$1)
                 )
               )
+              (get_local $$1)
             )
+            (get_local $$1)
           )
         )
       )
diff --git a/test/dot_s/store-results.wast b/test/dot_s/store-results.wast
index 190e18a92..c0d6ce905 100644
--- a/test/dot_s/store-results.wast
+++ b/test/dot_s/store-results.wast
@@ -34,14 +34,14 @@
             )
             (set_local $$1
               (i32.add
-                (i32.const 1)
                 (get_local $$1)
+                (i32.const 1)
               )
             )
             (br_if
               (i32.ne
-                (i32.const 256)
                 (get_local $$1)
+                (i32.const 256)
               )
               $BB1_1
             )
@@ -70,14 +70,14 @@
             )
             (set_local $$1
               (f32.add
-                (f32.const 1)
                 (get_local $$1)
+                (f32.const 1)
               )
             )
             (br_if
               (f32.ne
-                (f32.const 256)
                 (get_local $$1)
+                (f32.const 256)
               )
               $BB2_1
             )
diff --git a/test/dot_s/switch.wast b/test/dot_s/switch.wast
index 23d276092..0e7f8f03d 100644
--- a/test/dot_s/switch.wast
+++ b/test/dot_s/switch.wast
@@ -14,8 +14,8 @@
         (block $BB0_8
           (br_if
             (i32.gt_u
-              (i32.const 23)
               (get_local $$0)
+              (i32.const 23)
             )
             $BB0_8
           )
@@ -57,8 +57,8 @@
         (block $BB1_8
           (br_if
             (i64.gt_u
-              (i64.const 23)
               (get_local $$0)
+              (i64.const 23)
             )
             $BB1_8
           )
diff --git a/test/dot_s/varargs.wast b/test/dot_s/varargs.wast
index 3b455386e..c4e7e50b0 100644
--- a/test/dot_s/varargs.wast
+++ b/test/dot_s/varargs.wast
@@ -39,8 +39,8 @@
         (i32.store align=4
           (get_local $$0)
           (i32.add
-            (i32.const 4)
             (get_local $$1)
+            (i32.const 4)
           )
         )
         (br $fake_return_waka123
@@ -69,8 +69,8 @@
         (i32.store align=4
           (get_local $$0)
           (i32.add
-            (i32.const 4)
             (get_local $$1)
+            (i32.const 4)
           )
         )
         (br $fake_return_waka123
@@ -106,8 +106,8 @@
           (i32.store align=4
             (get_local $$1)
             (i32.add
-              (get_local $$3)
               (get_local $$2)
+              (get_local $$3)
             )
           )
         )
@@ -119,17 +119,17 @@
         (i32.store align=4
           (get_local $$1)
           (i32.add
-            (i32.const 16)
             (get_local $$2)
+            (i32.const 16)
           )
         )
         (i64.store align=8
-          (i64.load align=8
-            (get_local $$4)
-          )
           (i32.add
-            (get_local $$3)
             (get_local $$0)
+            (get_local $$3)
+          )
+          (i64.load align=8
+            (get_local $$4)
           )
         )
         (i64.store align=8