handle load in getMaxBits

3 files changed, 60 insertions, 0 deletions

diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp
index d28bf21a5..6dbb1b7e8 100644
--- a/src/passes/OptimizeInstructions.cpp
+++ b/src/passes/OptimizeInstructions.cpp
@@ -231,6 +231,8 @@ static Index getMaxBits(Expression* curr) {
   } else if (auto* set = curr->dynCast<SetLocal>()) {
     // a tee passes through the value
     return getMaxBits(set->value);
+  } else if (auto* load = curr->dynCast<Load>()) {
+    return 8 * load->bytes;
   }
   switch (curr->type) {
     case i32: return 32;
diff --git a/test/passes/optimize-instructions.txt b/test/passes/optimize-instructions.txt
index 9819cb8c7..01f1070f7 100644
--- a/test/passes/optimize-instructions.txt
+++ b/test/passes/optimize-instructions.txt
@@ -1218,4 +1218,24 @@
       )
     )
   )
+  (func $sign-ext-load (type $4) (param $0 i32) (param $1 i32)
+    (drop
+      (i32.load8_s
+        (i32.const 256)
+      )
+    )
+    (drop
+      (i32.shr_u
+        (i32.load8_s
+          (i32.const 256)
+        )
+        (i32.const 1)
+      )
+    )
+    (drop
+      (i32.load16_s
+        (i32.const 256)
+      )
+    )
+  )
 )
diff --git a/test/passes/optimize-instructions.wast b/test/passes/optimize-instructions.wast
index 1f0cd9094..9cdec0a6c 100644
--- a/test/passes/optimize-instructions.wast
+++ b/test/passes/optimize-instructions.wast
@@ -1525,4 +1525,42 @@
       )
     )
   )
+  (func $sign-ext-load (param $0 i32) (param $1 i32)
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (i32.load8_s ;; one byte, so perfect
+            (i32.const 256)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (i32.shr_u
+            (i32.load8_s ;; one byte, but reduced to 7
+              (i32.const 256)
+            )
+            (i32.const 1)
+          )
+          (i32.const 24)
+        )
+        (i32.const 24)
+      )
+    )
+    (drop
+      (i32.shr_s
+        (i32.shl
+          (i32.load16_s ;; two, so perfect
+            (i32.const 256)
+          )
+          (i32.const 16)
+        )
+        (i32.const 16)
+      )
+    )
+  )
 )