optimize loads+shifts into signed loads

8 files changed, 491 insertions, 428 deletions

diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp
index 2a5e7b8c4..003d2ff60 100644
--- a/src/passes/OptimizeInstructions.cpp
+++ b/src/passes/OptimizeInstructions.cpp
@@ -168,6 +168,12 @@ struct OptimizeInstructions : public WalkerPass<PostWalker<OptimizeInstructions,
   void visitExpression(Expression* curr) {
     // we may be able to apply multiple patterns, one may open opportunities that look deeper NB: patterns must not have cycles
     while (1) {
+      auto* handOptimized = handOptimize(curr);
+      if (handOptimized) {
+        curr = handOptimized;
+        replaceCurrent(curr);
+        continue;
+      }
       auto iter = database->patternMap.find(curr->_id);
       if (iter == database->patternMap.end()) return;
       auto& patterns = iter->second;
@@ -184,6 +190,27 @@ struct OptimizeInstructions : public WalkerPass<PostWalker<OptimizeInstructions,
       if (!more) break;
     }
   }
+
+  // Optimizations that don't yet fit in the pattern DSL, but could be eventually maybe
+  Expression* handOptimize(Expression* curr) {
+    if (auto* binary = curr->dynCast<Binary>()) {
+      // pattern match a load of 8 bits and a sign extend using a shl of 24 then shr_s of 24 as well, etc.
+      if (binary->op == BinaryOp::ShrSInt32 && binary->right->is<Const>()) {
+        auto shifts = binary->right->cast<Const>()->value.geti32();
+        if (shifts == 24 || shifts == 16) {
+          auto* left = binary->left->dynCast<Binary>();
+          if (left && left->op == ShlInt32 && left->right->is<Const>() && left->right->cast<Const>()->value.geti32() == shifts) {
+            auto* load = left->left->dynCast<Load>();
+            if (load && ((load->bytes == 1 && shifts == 24) || (load->bytes == 2 && shifts == 16))) {
+              load->signed_ = true;
+              return load;
+            }
+          }
+        }
+      }
+    }
+    return nullptr;
+  }
 };
 
 Pass *createOptimizeInstructionsPass() {
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
index 10da1e436..15b9aa473 100644
--- a/test/emcc_hello_world.fromasm
+++ b/test/emcc_hello_world.fromasm
@@ -467,14 +467,8 @@
                 )
               )
               (if
-                (i32.shr_s
-                  (i32.shl
-                    (i32.load8_s
-                      (get_local $2)
-                    )
-                    (i32.const 24)
-                  )
-                  (i32.const 24)
+                (i32.load8_s
+                  (get_local $2)
                 )
                 (set_local $2
                   (get_local $0)
@@ -1415,14 +1409,8 @@
           )
           (if
             (i32.lt_s
-              (i32.shr_s
-                (i32.shl
-                  (i32.load8_s offset=74
-                    (get_local $0)
-                  )
-                  (i32.const 24)
-                )
-                (i32.const 24)
+              (i32.load8_s offset=74
+                (get_local $0)
               )
               (i32.const 1)
             )
@@ -1697,14 +1685,8 @@
               (block $label$break$L10
                 (if
                   (i32.gt_s
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=75
-                          (get_local $2)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=75
+                      (get_local $2)
                     )
                     (i32.const -1)
                   )
@@ -1729,22 +1711,16 @@
                         )
                         (if
                           (i32.eq
-                            (i32.shr_s
-                              (i32.shl
-                                (i32.load8_s
+                            (i32.load8_s
+                              (i32.add
+                                (get_local $0)
+                                (tee_local $4
                                   (i32.add
-                                    (get_local $0)
-                                    (tee_local $4
-                                      (i32.add
-                                        (get_local $3)
-                                        (i32.const -1)
-                                      )
-                                    )
+                                    (get_local $3)
+                                    (i32.const -1)
                                   )
                                 )
-                                (i32.const 24)
                               )
-                              (i32.const 24)
                             )
                             (i32.const 10)
                           )
@@ -1840,19 +1816,13 @@
         (i32.or
           (i32.add
             (tee_local $1
-              (i32.shr_s
-                (i32.shl
-                  (i32.load8_s
-                    (tee_local $2
-                      (i32.add
-                        (get_local $0)
-                        (i32.const 74)
-                      )
-                    )
+              (i32.load8_s
+                (tee_local $2
+                  (i32.add
+                    (get_local $0)
+                    (i32.const 74)
                   )
-                  (i32.const 24)
                 )
-                (i32.const 24)
               )
             )
             (i32.const 255)
@@ -2189,14 +2159,8 @@
             (block $while-out$1
               (if
                 (i32.eq
-                  (i32.shr_s
-                    (i32.shl
-                      (i32.load8_s
-                        (get_local $2)
-                      )
-                      (i32.const 24)
-                    )
-                    (i32.const 24)
+                  (i32.load8_s
+                    (get_local $2)
                   )
                   (i32.shr_s
                     (i32.shl
@@ -2320,14 +2284,8 @@
         )
         (if
           (i32.eq
-            (i32.shr_s
-              (i32.shl
-                (i32.load8_s
-                  (get_local $5)
-                )
-                (i32.const 24)
-              )
-              (i32.const 24)
+            (i32.load8_s
+              (get_local $5)
             )
             (i32.shr_s
               (i32.shl
@@ -2476,14 +2434,8 @@
               (block $while-out$7
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (get_local $9)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s
+                      (get_local $9)
                     )
                     (i32.shr_s
                       (i32.shl
@@ -3066,14 +3018,8 @@
                 )
                 (if
                   (i32.ne
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=1
-                          (get_local $54)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=1
+                      (get_local $54)
                     )
                     (i32.const 37)
                   )
@@ -3095,19 +3041,13 @@
                 )
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (tee_local $1
-                            (i32.add
-                              (get_local $54)
-                              (i32.const 2)
-                            )
-                          )
+                    (i32.load8_s
+                      (tee_local $1
+                        (i32.add
+                          (get_local $54)
+                          (i32.const 2)
                         )
-                        (i32.const 24)
                       )
-                      (i32.const 24)
                     )
                     (i32.const 37)
                   )
@@ -3211,14 +3151,8 @@
                       (get_local $5)
                       (tee_local $7
                         (i32.eq
-                          (i32.shr_s
-                            (i32.shl
-                              (i32.load8_s offset=2
-                                (get_local $41)
-                              )
-                              (i32.const 24)
-                            )
-                            (i32.const 24)
+                          (i32.load8_s offset=2
+                            (get_local $41)
                           )
                           (i32.const 36)
                         )
@@ -3370,19 +3304,13 @@
                 (i32.lt_u
                   (tee_local $1
                     (i32.add
-                      (i32.shr_s
-                        (i32.shl
-                          (i32.load8_s
-                            (tee_local $6
-                              (i32.add
-                                (get_local $9)
-                                (i32.const 1)
-                              )
-                            )
+                      (i32.load8_s
+                        (tee_local $6
+                          (i32.add
+                            (get_local $9)
+                            (i32.const 1)
                           )
-                          (i32.const 24)
                         )
-                        (i32.const 24)
                       )
                       (i32.const -48)
                     )
@@ -3391,14 +3319,8 @@
                 )
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=2
-                          (get_local $9)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=2
+                      (get_local $9)
                     )
                     (i32.const 36)
                   )
@@ -3419,14 +3341,8 @@
                           (get_local $3)
                           (i32.shl
                             (i32.add
-                              (i32.shr_s
-                                (i32.shl
-                                  (i32.load8_s
-                                    (get_local $6)
-                                  )
-                                  (i32.const 24)
-                                )
-                                (i32.const 24)
+                              (i32.load8_s
+                                (get_local $6)
                               )
                               (i32.const -48)
                             )
@@ -3601,19 +3517,13 @@
                       (i32.ge_u
                         (tee_local $6
                           (i32.add
-                            (i32.shr_s
-                              (i32.shl
-                                (i32.load8_s
-                                  (tee_local $1
-                                    (i32.add
-                                      (get_local $1)
-                                      (i32.const 1)
-                                    )
-                                  )
+                            (i32.load8_s
+                              (tee_local $1
+                                (i32.add
+                                  (get_local $1)
+                                  (i32.const 1)
                                 )
-                                (i32.const 24)
                               )
-                              (i32.const 24)
                             )
                             (i32.const -48)
                           )
@@ -3663,14 +3573,8 @@
           (block $label$break$L46
             (if
               (i32.eq
-                (i32.shr_s
-                  (i32.shl
-                    (i32.load8_s
-                      (get_local $9)
-                    )
-                    (i32.const 24)
-                  )
-                  (i32.const 24)
+                (i32.load8_s
+                  (get_local $9)
                 )
                 (i32.const 46)
               )
@@ -3743,19 +3647,13 @@
                         (i32.ge_u
                           (tee_local $6
                             (i32.add
-                              (i32.shr_s
-                                (i32.shl
-                                  (i32.load8_s
-                                    (tee_local $1
-                                      (i32.add
-                                        (get_local $1)
-                                        (i32.const 1)
-                                      )
-                                    )
+                              (i32.load8_s
+                                (tee_local $1
+                                  (i32.add
+                                    (get_local $1)
+                                    (i32.const 1)
                                   )
-                                  (i32.const 24)
                                 )
-                                (i32.const 24)
                               )
                               (i32.const -48)
                             )
@@ -3779,19 +3677,13 @@
                   (i32.lt_u
                     (tee_local $1
                       (i32.add
-                        (i32.shr_s
-                          (i32.shl
-                            (i32.load8_s
-                              (tee_local $6
-                                (i32.add
-                                  (get_local $9)
-                                  (i32.const 2)
-                                )
-                              )
+                        (i32.load8_s
+                          (tee_local $6
+                            (i32.add
+                              (get_local $9)
+                              (i32.const 2)
                             )
-                            (i32.const 24)
                           )
-                          (i32.const 24)
                         )
                         (i32.const -48)
                       )
@@ -3800,14 +3692,8 @@
                   )
                   (if
                     (i32.eq
-                      (i32.shr_s
-                        (i32.shl
-                          (i32.load8_s offset=3
-                            (get_local $9)
-                          )
-                          (i32.const 24)
-                        )
-                        (i32.const 24)
+                      (i32.load8_s offset=3
+                        (get_local $9)
                       )
                       (i32.const 36)
                     )
@@ -3828,14 +3714,8 @@
                             (get_local $3)
                             (i32.shl
                               (i32.add
-                                (i32.shr_s
-                                  (i32.shl
-                                    (i32.load8_s
-                                      (get_local $6)
-                                    )
-                                    (i32.const 24)
-                                  )
-                                  (i32.const 24)
+                                (i32.load8_s
+                                  (get_local $6)
                                 )
                                 (i32.const -48)
                               )
@@ -3921,14 +3801,8 @@
               (i32.gt_u
                 (tee_local $1
                   (i32.add
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (get_local $11)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s
+                      (get_local $11)
                     )
                     (i32.const -65)
                   )
@@ -4143,14 +4017,8 @@
             (i32.eq
               (i32.and
                 (tee_local $1
-                  (i32.shr_s
-                    (i32.shl
-                      (i32.load8_s
-                        (get_local $11)
-                      )
-                      (i32.const 24)
-                    )
-                    (i32.const 24)
+                  (i32.load8_s
+                    (get_local $11)
                   )
                 )
                 (i32.const 15)
@@ -5023,14 +4891,8 @@
                                   (get_local $30)
                                 )
                                 (i32.eq
-                                  (i32.shr_s
-                                    (i32.shl
-                                      (i32.load8_s
-                                        (get_local $9)
-                                      )
-                                      (i32.const 24)
-                                    )
-                                    (i32.const 24)
+                                  (i32.load8_s
+                                    (get_local $9)
                                   )
                                   (i32.const 45)
                                 )
@@ -6182,14 +6044,8 @@
                                         (br_if $do-once$90
                                           (get_local $14)
                                           (i32.ne
-                                            (i32.shr_s
-                                              (i32.shl
-                                                (i32.load8_s
-                                                  (get_local $39)
-                                                )
-                                                (i32.const 24)
-                                              )
-                                              (i32.const 24)
+                                            (i32.load8_s
+                                              (get_local $39)
                                             )
                                             (i32.const 45)
                                           )
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
index bcdebdd99..9d70778d5 100644
--- a/test/emcc_hello_world.fromasm.imprecise
+++ b/test/emcc_hello_world.fromasm.imprecise
@@ -460,14 +460,8 @@
                 )
               )
               (if
-                (i32.shr_s
-                  (i32.shl
-                    (i32.load8_s
-                      (get_local $2)
-                    )
-                    (i32.const 24)
-                  )
-                  (i32.const 24)
+                (i32.load8_s
+                  (get_local $2)
                 )
                 (set_local $2
                   (get_local $0)
@@ -1408,14 +1402,8 @@
           )
           (if
             (i32.lt_s
-              (i32.shr_s
-                (i32.shl
-                  (i32.load8_s offset=74
-                    (get_local $0)
-                  )
-                  (i32.const 24)
-                )
-                (i32.const 24)
+              (i32.load8_s offset=74
+                (get_local $0)
               )
               (i32.const 1)
             )
@@ -1690,14 +1678,8 @@
               (block $label$break$L10
                 (if
                   (i32.gt_s
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=75
-                          (get_local $2)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=75
+                      (get_local $2)
                     )
                     (i32.const -1)
                   )
@@ -1722,22 +1704,16 @@
                         )
                         (if
                           (i32.eq
-                            (i32.shr_s
-                              (i32.shl
-                                (i32.load8_s
+                            (i32.load8_s
+                              (i32.add
+                                (get_local $0)
+                                (tee_local $4
                                   (i32.add
-                                    (get_local $0)
-                                    (tee_local $4
-                                      (i32.add
-                                        (get_local $3)
-                                        (i32.const -1)
-                                      )
-                                    )
+                                    (get_local $3)
+                                    (i32.const -1)
                                   )
                                 )
-                                (i32.const 24)
                               )
-                              (i32.const 24)
                             )
                             (i32.const 10)
                           )
@@ -1833,19 +1809,13 @@
         (i32.or
           (i32.add
             (tee_local $1
-              (i32.shr_s
-                (i32.shl
-                  (i32.load8_s
-                    (tee_local $2
-                      (i32.add
-                        (get_local $0)
-                        (i32.const 74)
-                      )
-                    )
+              (i32.load8_s
+                (tee_local $2
+                  (i32.add
+                    (get_local $0)
+                    (i32.const 74)
                   )
-                  (i32.const 24)
                 )
-                (i32.const 24)
               )
             )
             (i32.const 255)
@@ -2182,14 +2152,8 @@
             (block $while-out$1
               (if
                 (i32.eq
-                  (i32.shr_s
-                    (i32.shl
-                      (i32.load8_s
-                        (get_local $2)
-                      )
-                      (i32.const 24)
-                    )
-                    (i32.const 24)
+                  (i32.load8_s
+                    (get_local $2)
                   )
                   (i32.shr_s
                     (i32.shl
@@ -2313,14 +2277,8 @@
         )
         (if
           (i32.eq
-            (i32.shr_s
-              (i32.shl
-                (i32.load8_s
-                  (get_local $5)
-                )
-                (i32.const 24)
-              )
-              (i32.const 24)
+            (i32.load8_s
+              (get_local $5)
             )
             (i32.shr_s
               (i32.shl
@@ -2469,14 +2427,8 @@
               (block $while-out$7
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (get_local $9)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s
+                      (get_local $9)
                     )
                     (i32.shr_s
                       (i32.shl
@@ -3059,14 +3011,8 @@
                 )
                 (if
                   (i32.ne
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=1
-                          (get_local $54)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=1
+                      (get_local $54)
                     )
                     (i32.const 37)
                   )
@@ -3088,19 +3034,13 @@
                 )
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (tee_local $1
-                            (i32.add
-                              (get_local $54)
-                              (i32.const 2)
-                            )
-                          )
+                    (i32.load8_s
+                      (tee_local $1
+                        (i32.add
+                          (get_local $54)
+                          (i32.const 2)
                         )
-                        (i32.const 24)
                       )
-                      (i32.const 24)
                     )
                     (i32.const 37)
                   )
@@ -3204,14 +3144,8 @@
                       (get_local $5)
                       (tee_local $7
                         (i32.eq
-                          (i32.shr_s
-                            (i32.shl
-                              (i32.load8_s offset=2
-                                (get_local $41)
-                              )
-                              (i32.const 24)
-                            )
-                            (i32.const 24)
+                          (i32.load8_s offset=2
+                            (get_local $41)
                           )
                           (i32.const 36)
                         )
@@ -3363,19 +3297,13 @@
                 (i32.lt_u
                   (tee_local $1
                     (i32.add
-                      (i32.shr_s
-                        (i32.shl
-                          (i32.load8_s
-                            (tee_local $6
-                              (i32.add
-                                (get_local $9)
-                                (i32.const 1)
-                              )
-                            )
+                      (i32.load8_s
+                        (tee_local $6
+                          (i32.add
+                            (get_local $9)
+                            (i32.const 1)
                           )
-                          (i32.const 24)
                         )
-                        (i32.const 24)
                       )
                       (i32.const -48)
                     )
@@ -3384,14 +3312,8 @@
                 )
                 (if
                   (i32.eq
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s offset=2
-                          (get_local $9)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s offset=2
+                      (get_local $9)
                     )
                     (i32.const 36)
                   )
@@ -3412,14 +3334,8 @@
                           (get_local $3)
                           (i32.shl
                             (i32.add
-                              (i32.shr_s
-                                (i32.shl
-                                  (i32.load8_s
-                                    (get_local $6)
-                                  )
-                                  (i32.const 24)
-                                )
-                                (i32.const 24)
+                              (i32.load8_s
+                                (get_local $6)
                               )
                               (i32.const -48)
                             )
@@ -3594,19 +3510,13 @@
                       (i32.ge_u
                         (tee_local $6
                           (i32.add
-                            (i32.shr_s
-                              (i32.shl
-                                (i32.load8_s
-                                  (tee_local $1
-                                    (i32.add
-                                      (get_local $1)
-                                      (i32.const 1)
-                                    )
-                                  )
+                            (i32.load8_s
+                              (tee_local $1
+                                (i32.add
+                                  (get_local $1)
+                                  (i32.const 1)
                                 )
-                                (i32.const 24)
                               )
-                              (i32.const 24)
                             )
                             (i32.const -48)
                           )
@@ -3656,14 +3566,8 @@
           (block $label$break$L46
             (if
               (i32.eq
-                (i32.shr_s
-                  (i32.shl
-                    (i32.load8_s
-                      (get_local $9)
-                    )
-                    (i32.const 24)
-                  )
-                  (i32.const 24)
+                (i32.load8_s
+                  (get_local $9)
                 )
                 (i32.const 46)
               )
@@ -3736,19 +3640,13 @@
                         (i32.ge_u
                           (tee_local $6
                             (i32.add
-                              (i32.shr_s
-                                (i32.shl
-                                  (i32.load8_s
-                                    (tee_local $1
-                                      (i32.add
-                                        (get_local $1)
-                                        (i32.const 1)
-                                      )
-                                    )
+                              (i32.load8_s
+                                (tee_local $1
+                                  (i32.add
+                                    (get_local $1)
+                                    (i32.const 1)
                                   )
-                                  (i32.const 24)
                                 )
-                                (i32.const 24)
                               )
                               (i32.const -48)
                             )
@@ -3772,19 +3670,13 @@
                   (i32.lt_u
                     (tee_local $1
                       (i32.add
-                        (i32.shr_s
-                          (i32.shl
-                            (i32.load8_s
-                              (tee_local $6
-                                (i32.add
-                                  (get_local $9)
-                                  (i32.const 2)
-                                )
-                              )
+                        (i32.load8_s
+                          (tee_local $6
+                            (i32.add
+                              (get_local $9)
+                              (i32.const 2)
                             )
-                            (i32.const 24)
                           )
-                          (i32.const 24)
                         )
                         (i32.const -48)
                       )
@@ -3793,14 +3685,8 @@
                   )
                   (if
                     (i32.eq
-                      (i32.shr_s
-                        (i32.shl
-                          (i32.load8_s offset=3
-                            (get_local $9)
-                          )
-                          (i32.const 24)
-                        )
-                        (i32.const 24)
+                      (i32.load8_s offset=3
+                        (get_local $9)
                       )
                       (i32.const 36)
                     )
@@ -3821,14 +3707,8 @@
                             (get_local $3)
                             (i32.shl
                               (i32.add
-                                (i32.shr_s
-                                  (i32.shl
-                                    (i32.load8_s
-                                      (get_local $6)
-                                    )
-                                    (i32.const 24)
-                                  )
-                                  (i32.const 24)
+                                (i32.load8_s
+                                  (get_local $6)
                                 )
                                 (i32.const -48)
                               )
@@ -3914,14 +3794,8 @@
               (i32.gt_u
                 (tee_local $1
                   (i32.add
-                    (i32.shr_s
-                      (i32.shl
-                        (i32.load8_s
-                          (get_local $11)
-                        )
-                        (i32.const 24)
-                      )
-                      (i32.const 24)
+                    (i32.load8_s
+                      (get_local $11)
                     )
                     (i32.const -65)
                   )
@@ -4136,14 +4010,8 @@
             (i32.eq
               (i32.and
                 (tee_local $1
-                  (i32.shr_s
-                    (i32.shl
-                      (i32.load8_s
-                        (get_local $11)
-                      )
-                      (i32.const 24)
-                    )
-                    (i32.const 24)
+                  (i32.load8_s
+                    (get_local $11)
                   )
                 )
                 (i32.const 15)
@@ -5016,14 +4884,8 @@
                                   (get_local $30)
                                 )
                                 (i32.eq
-                                  (i32.shr_s
-                                    (i32.shl
-                                      (i32.load8_s
-                                        (get_local $9)
-                                      )
-                                      (i32.const 24)
-                                    )
-                                    (i32.const 24)
+                                  (i32.load8_s
+                                    (get_local $9)
                                   )
                                   (i32.const 45)
                                 )
@@ -6175,14 +6037,8 @@
                                         (br_if $do-once$90
                                           (get_local $14)
                                           (i32.ne
-                                            (i32.shr_s
-                                              (i32.shl
-                                                (i32.load8_s
-                                                  (get_local $39)
-                                                )
-                                                (i32.const 24)
-                                              )
-                                              (i32.const 24)
+                                            (i32.load8_s
+                                              (get_local $39)
                                             )
                                             (i32.const 45)
                                           )
diff --git a/test/unit.asm.js b/test/unit.asm.js
index e76ae1ef9..a25e60721 100644
--- a/test/unit.asm.js
+++ b/test/unit.asm.js
@@ -330,6 +330,18 @@ function asm(global, env, buffer) {
     y = 3 ? +abort(7) : 4.5;
   }
 
+  function loadSigned(x) {
+    x = x | 0;
+    loadSigned(HEAP8[x >> 0] << 24 >> 24);
+    loadSigned(HEAPU8[x >> 0] << 24 >> 24);
+    loadSigned(HEAP16[x >> 1] << 16 >> 16);
+    loadSigned(HEAPU16[x >> 1] << 16 >> 16);
+    loadSigned(HEAP8[x >> 0] << 24 >> 16);
+    loadSigned(HEAPU8[x >> 0] << 16 >> 24);
+    loadSigned(HEAP16[x >> 1] << 16 >> 24);
+    loadSigned(HEAPU16[x >> 1] << 24 >> 16);
+  }
+
   function z() {
   }
   function w() {
diff --git a/test/unit.fromasm b/test/unit.fromasm
index 33dd153af..bafc17f3f 100644
--- a/test/unit.fromasm
+++ b/test/unit.fromasm
@@ -616,4 +616,70 @@
       )
     )
   )
+  (func $loadSigned (param $0 i32)
+    (call $loadSigned
+      (i32.load8_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load8_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load16_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load16_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $0)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $0)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $0)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $0)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+  )
 )
diff --git a/test/unit.fromasm.imprecise b/test/unit.fromasm.imprecise
index 35b8b1768..cdaca7e10 100644
--- a/test/unit.fromasm.imprecise
+++ b/test/unit.fromasm.imprecise
@@ -597,4 +597,70 @@
       )
     )
   )
+  (func $loadSigned (param $0 i32)
+    (call $loadSigned
+      (i32.load8_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load8_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load16_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.load16_s
+        (get_local $0)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $0)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $0)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $0)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $0)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+  )
 )
diff --git a/test/unit.fromasm.imprecise.no-opts b/test/unit.fromasm.imprecise.no-opts
index c41075df0..244ef7cd8 100644
--- a/test/unit.fromasm.imprecise.no-opts
+++ b/test/unit.fromasm.imprecise.no-opts
@@ -1001,6 +1001,96 @@
       )
     )
   )
+  (func $loadSigned (param $x i32)
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+  )
   (func $z
     (nop)
   )
diff --git a/test/unit.fromasm.no-opts b/test/unit.fromasm.no-opts
index 2162d9ed7..1c0e87f6a 100644
--- a/test/unit.fromasm.no-opts
+++ b/test/unit.fromasm.no-opts
@@ -1007,6 +1007,96 @@
       )
     )
   )
+  (func $loadSigned (param $x i32)
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_u
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s
+            (get_local $x)
+          )
+          (i32.const 16)
+        )
+        (i32.const 24)
+      )
+    )
+    (call $loadSigned
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_u
+            (get_local $x)
+          )
+          (i32.const 24)
+        )
+        (i32.const 16)
+      )
+    )
+  )
   (func $z
     (nop)
   )