take into account loads into local info in OptimizeInstructions

5 files changed, 184 insertions, 97 deletions

diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp
index 8e5ff6c12..026f3f7ec 100644
--- a/src/passes/OptimizeInstructions.cpp
+++ b/src/passes/OptimizeInstructions.cpp
@@ -386,15 +386,18 @@ struct LocalScanner : PostWalker<LocalScanner, Visitor<LocalScanner>> {
     // an integer var, worth processing
     auto& info = localInfo[curr->index];
     info.maxBits = std::max(info.maxBits, getMaxBits(curr->value, this));
+    auto signExtBits = LocalInfo::kUnknown;
     if (getSignExt(curr->value)) {
-      auto bits = getSignExtBits(curr->value);
-      if (info.signExtedBits == 0) {
-        info.signExtedBits = bits; // first info we see
-      } else if (info.signExtedBits != bits) {
-        info.signExtedBits = LocalInfo::kUnknown; // contradictory information, give up
+      signExtBits = getSignExtBits(curr->value);
+    } else if (auto* load = getFallthroughDynCast<Load>(curr->value)) {
+      if (load->signed_) {
+        signExtBits = load->bytes * 8;
       }
-    } else {
-      info.signExtedBits = LocalInfo::kUnknown; // an input which isn't even a sign ext, give up
+    }
+    if (info.signExtedBits == 0) {
+      info.signExtedBits = signExtBits; // first info we see
+    } else if (info.signExtedBits != signExtBits) {
+      info.signExtedBits = LocalInfo::kUnknown; // contradictory information, give up
     }
   }
 
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
index f01ba2f5d..c314d887c 100644
--- a/test/emcc_hello_world.fromasm
+++ b/test/emcc_hello_world.fromasm
@@ -325,7 +325,7 @@
             )
           )
           (if
-            (i32.load8_u
+            (i32.load8_s
               (get_local $0)
             )
             (block
@@ -1468,7 +1468,7 @@
                 )
                 (if
                   (i32.ne
-                    (i32.load8_u
+                    (i32.load8_s
                       (i32.add
                         (get_local $0)
                         (tee_local $6
@@ -1864,12 +1864,15 @@
               (loop $while-in
                 (if
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (get_local $2)
                     )
-                    (i32.and
-                      (get_local $4)
-                      (i32.const 255)
+                    (i32.shr_s
+                      (i32.shl
+                        (get_local $4)
+                        (i32.const 24)
+                      )
+                      (i32.const 24)
                     )
                   )
                   (block
@@ -1938,14 +1941,20 @@
       )
       (if
         (i32.ne
-          (i32.load8_u
+          (i32.load8_s
             (get_local $2)
           )
-          (tee_local $1
-            (i32.and
-              (get_local $1)
-              (i32.const 255)
+          (i32.shr_s
+            (i32.shl
+              (tee_local $1
+                (i32.and
+                  (get_local $1)
+                  (i32.const 255)
+                )
+              )
+              (i32.const 24)
             )
+            (i32.const 24)
           )
         )
         (block
@@ -2026,12 +2035,15 @@
           (loop $while-in5
             (br_if $label$break$L8
               (i32.eq
-                (i32.load8_u
+                (i32.load8_s
                   (get_local $2)
                 )
-                (i32.and
-                  (get_local $1)
-                  (i32.const 255)
+                (i32.shr_s
+                  (i32.shl
+                    (get_local $1)
+                    (i32.const 24)
+                  )
+                  (i32.const 24)
                 )
               )
             )
@@ -2498,13 +2510,10 @@
             )
             (br_if $__rjti$9
               (i32.eqz
-                (i32.and
-                  (tee_local $7
-                    (i32.load8_s
-                      (get_local $5)
-                    )
+                (tee_local $7
+                  (i32.load8_s
+                    (get_local $5)
                   )
-                  (i32.const 255)
                 )
               )
             )
@@ -2556,7 +2565,7 @@
               (loop $while-in
                 (br_if $label$break$L12
                   (i32.ne
-                    (i32.load8_u offset=1
+                    (i32.load8_s offset=1
                       (get_local $6)
                     )
                     (i32.const 37)
@@ -2570,7 +2579,7 @@
                 )
                 (br_if $while-in
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (tee_local $6
                         (i32.add
                           (get_local $6)
@@ -2656,7 +2665,7 @@
                           (get_local $10)
                           (tee_local $11
                             (i32.eq
-                              (i32.load8_u offset=2
+                              (i32.load8_s offset=2
                                 (get_local $6)
                               )
                               (i32.const 36)
@@ -2825,7 +2834,7 @@
                         )
                         (br_if $__rjti$0
                           (i32.ne
-                            (i32.load8_u offset=2
+                            (i32.load8_s offset=2
                               (get_local $10)
                             )
                             (i32.const 36)
@@ -3057,7 +3066,7 @@
               (set_local $6
                 (if i32
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (get_local $10)
                     )
                     (i32.const 46)
@@ -3065,18 +3074,15 @@
                   (block i32
                     (if
                       (i32.ne
-                        (i32.and
-                          (tee_local $8
-                            (i32.load8_s
-                              (tee_local $6
-                                (i32.add
-                                  (get_local $10)
-                                  (i32.const 1)
-                                )
+                        (tee_local $8
+                          (i32.load8_s
+                            (tee_local $6
+                              (i32.add
+                                (get_local $10)
+                                (i32.const 1)
                               )
                             )
                           )
-                          (i32.const 255)
                         )
                         (i32.const 42)
                       )
@@ -3085,13 +3091,7 @@
                           (i32.lt_u
                             (tee_local $9
                               (i32.add
-                                (i32.shr_s
-                                  (i32.shl
-                                    (get_local $8)
-                                    (i32.const 24)
-                                  )
-                                  (i32.const 24)
-                                )
+                                (get_local $8)
                                 (i32.const -48)
                               )
                             )
@@ -3175,7 +3175,7 @@
                       )
                       (if
                         (i32.eq
-                          (i32.load8_u offset=3
+                          (i32.load8_s offset=3
                             (get_local $10)
                           )
                           (i32.const 36)
@@ -4201,7 +4201,7 @@
                                               )
                                               (if f64
                                                 (i32.eq
-                                                  (i32.load8_u
+                                                  (i32.load8_s
                                                     (get_local $9)
                                                   )
                                                   (i32.const 45)
@@ -5261,7 +5261,7 @@
                                                     (block
                                                       (br_if $do-once83
                                                         (i32.ne
-                                                          (i32.load8_u
+                                                          (i32.load8_s
                                                             (get_local $31)
                                                           )
                                                           (i32.const 45)
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
index 22878e91f..1fa81ba43 100644
--- a/test/emcc_hello_world.fromasm.imprecise
+++ b/test/emcc_hello_world.fromasm.imprecise
@@ -322,7 +322,7 @@
             )
           )
           (if
-            (i32.load8_u
+            (i32.load8_s
               (get_local $0)
             )
             (block
@@ -1465,7 +1465,7 @@
                 )
                 (if
                   (i32.ne
-                    (i32.load8_u
+                    (i32.load8_s
                       (i32.add
                         (get_local $0)
                         (tee_local $6
@@ -1861,12 +1861,15 @@
               (loop $while-in
                 (if
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (get_local $2)
                     )
-                    (i32.and
-                      (get_local $4)
-                      (i32.const 255)
+                    (i32.shr_s
+                      (i32.shl
+                        (get_local $4)
+                        (i32.const 24)
+                      )
+                      (i32.const 24)
                     )
                   )
                   (block
@@ -1935,14 +1938,20 @@
       )
       (if
         (i32.ne
-          (i32.load8_u
+          (i32.load8_s
             (get_local $2)
           )
-          (tee_local $1
-            (i32.and
-              (get_local $1)
-              (i32.const 255)
+          (i32.shr_s
+            (i32.shl
+              (tee_local $1
+                (i32.and
+                  (get_local $1)
+                  (i32.const 255)
+                )
+              )
+              (i32.const 24)
             )
+            (i32.const 24)
           )
         )
         (block
@@ -2023,12 +2032,15 @@
           (loop $while-in5
             (br_if $label$break$L8
               (i32.eq
-                (i32.load8_u
+                (i32.load8_s
                   (get_local $2)
                 )
-                (i32.and
-                  (get_local $1)
-                  (i32.const 255)
+                (i32.shr_s
+                  (i32.shl
+                    (get_local $1)
+                    (i32.const 24)
+                  )
+                  (i32.const 24)
                 )
               )
             )
@@ -2441,13 +2453,10 @@
             )
             (br_if $__rjti$9
               (i32.eqz
-                (i32.and
-                  (tee_local $7
-                    (i32.load8_s
-                      (get_local $5)
-                    )
+                (tee_local $7
+                  (i32.load8_s
+                    (get_local $5)
                   )
-                  (i32.const 255)
                 )
               )
             )
@@ -2499,7 +2508,7 @@
               (loop $while-in
                 (br_if $label$break$L12
                   (i32.ne
-                    (i32.load8_u offset=1
+                    (i32.load8_s offset=1
                       (get_local $6)
                     )
                     (i32.const 37)
@@ -2513,7 +2522,7 @@
                 )
                 (br_if $while-in
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (tee_local $6
                         (i32.add
                           (get_local $6)
@@ -2599,7 +2608,7 @@
                           (get_local $10)
                           (tee_local $11
                             (i32.eq
-                              (i32.load8_u offset=2
+                              (i32.load8_s offset=2
                                 (get_local $6)
                               )
                               (i32.const 36)
@@ -2768,7 +2777,7 @@
                         )
                         (br_if $__rjti$0
                           (i32.ne
-                            (i32.load8_u offset=2
+                            (i32.load8_s offset=2
                               (get_local $10)
                             )
                             (i32.const 36)
@@ -3000,7 +3009,7 @@
               (set_local $6
                 (if i32
                   (i32.eq
-                    (i32.load8_u
+                    (i32.load8_s
                       (get_local $10)
                     )
                     (i32.const 46)
@@ -3008,18 +3017,15 @@
                   (block i32
                     (if
                       (i32.ne
-                        (i32.and
-                          (tee_local $8
-                            (i32.load8_s
-                              (tee_local $6
-                                (i32.add
-                                  (get_local $10)
-                                  (i32.const 1)
-                                )
+                        (tee_local $8
+                          (i32.load8_s
+                            (tee_local $6
+                              (i32.add
+                                (get_local $10)
+                                (i32.const 1)
                               )
                             )
                           )
-                          (i32.const 255)
                         )
                         (i32.const 42)
                       )
@@ -3028,13 +3034,7 @@
                           (i32.lt_u
                             (tee_local $9
                               (i32.add
-                                (i32.shr_s
-                                  (i32.shl
-                                    (get_local $8)
-                                    (i32.const 24)
-                                  )
-                                  (i32.const 24)
-                                )
+                                (get_local $8)
                                 (i32.const -48)
                               )
                             )
@@ -3118,7 +3118,7 @@
                       )
                       (if
                         (i32.eq
-                          (i32.load8_u offset=3
+                          (i32.load8_s offset=3
                             (get_local $10)
                           )
                           (i32.const 36)
@@ -4162,7 +4162,7 @@
                                                   (get_local $15)
                                                 )
                                                 (i32.eq
-                                                  (i32.load8_u
+                                                  (i32.load8_s
                                                     (get_local $9)
                                                   )
                                                   (i32.const 45)
@@ -5198,7 +5198,7 @@
                                                     (block
                                                       (br_if $do-once83
                                                         (i32.ne
-                                                          (i32.load8_u
+                                                          (i32.load8_s
                                                             (get_local $31)
                                                           )
                                                           (i32.const 45)
diff --git a/test/passes/optimize-instructions.txt b/test/passes/optimize-instructions.txt
index 80d264e94..550859185 100644
--- a/test/passes/optimize-instructions.txt
+++ b/test/passes/optimize-instructions.txt
@@ -1601,4 +1601,46 @@
       )
     )
   )
+  (func $local-info-sign-ext-already-exted-by-load (type $4) (param $0 i32) (param $1 i32)
+    (local $x i32)
+    (local $y i32)
+    (local $z i32)
+    (local $w i32)
+    (set_local $x
+      (i32.load8_s
+        (i32.const 1024)
+      )
+    )
+    (drop
+      (get_local $x)
+    )
+    (set_local $y
+      (i32.load8_u
+        (i32.const 1024)
+      )
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (get_local $y)
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (set_local $z
+      (i32.load16_s
+        (i32.const 1024)
+      )
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (get_local $z)
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+  )
 )
diff --git a/test/passes/optimize-instructions.wast b/test/passes/optimize-instructions.wast
index 0169bb054..2f2333867 100644
--- a/test/passes/optimize-instructions.wast
+++ b/test/passes/optimize-instructions.wast
@@ -1978,4 +1978,46 @@
       )
     )
   )
+  (func $local-info-sign-ext-already-exted-by-load (param $0 i32) (param $1 i32)
+    (local $x i32)
+    (local $y i32)
+    (local $z i32)
+    (local $w i32)
+    (set_local $x
+      (i32.load8_s (i32.const 1024)) ;; 8 bits, sign extended, no need to do it again
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (get_local $x)
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (set_local $y
+      (i32.load8_u (i32.const 1024)) ;; 8 bits, zext, so bad
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (get_local $y)
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (set_local $z
+      (i32.load16_s (i32.const 1024)) ;; 16 bits sign-extended, wrong size
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (get_local $z)
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+  )
 )