Remove SMI length check from Builtins::Generate_ArrayIndexOf

src/builtins/builtins-array.cc

Index: src/builtins/builtins-array.cc
diff --git a/src/builtins/builtins-array.cc b/src/builtins/builtins-array.cc
index e8ee8640c665b16a380fb3f3a2c65eabca1c8929..b0dfae667a809beae748bff82c0caa9247c112d7 100644
--- a/src/builtins/builtins-array.cc
+++ b/src/builtins/builtins-array.cc
@@ -2021,11 +2021,12 @@ void Builtins::Generate_ArrayIndexOf(compiler::CodeAssemblerState* state) {
 
   assembler.Bind(&init_len);
   {
-    // Handle case where JSArray length is not an Smi in the runtime
-    Node* len = assembler.LoadObjectField(array, JSArray::kLengthOffset);
-    assembler.GotoIfNot(assembler.TaggedIsSmi(len), &call_runtime);
+    // JSArray length is always an Smi for fast arrays.
+    Node* len = assembler.LoadAndUntagToWord32ObjectField(
+        array, JSArray::kLengthOffset);
+    CSA_ASSERT(&assembler, assembler.TaggedIsSmi(len));

[caitp, 2017/02/24 16:09:16]

You're doing a LoadAndUntagToWord32ObjectField, which untags the Smi.

You'd need to change this to just LoadObjectField() and untag it after the
assert, or otherwise load the object field without untagging in the assertion.

[caitp, 2017/02/24 16:12:42]

There may not be any point in untagging though, I think you could just modify
the entire stub to use Smi comparison and increment operations instead.

[Benedikt Meurer, 2017/02/24 17:50:06]

That's less efficient on 64-bit platforms, since the Smis cannot be immediates
there. So the fix here is to use:

Node* len = LoadAndUntagObjectField(...);
CSA_ASSERT(TaggedIsSmi(LoadObjectField(...));

[caitp, 2017/02/24 18:19:05]

Fair enough --- might be worth updating the ArrayIncludes fast path to untag the
index/length as well, then

 
-    len_var.Bind(assembler.SmiToWord(len));
+    len_var.Bind(len);
     assembler.Branch(assembler.WordEqual(len_var.value(), intptr_zero),
                      &return_not_found, &init_k);
   }