[string] Optimize ToDirectStringAssembler for sequential strings

ToDirectStringAssembler is used in StringCharCodeAt and SubString
(which uses StringCharCodeAt internally). SubString is used all over the
place (e.g. RegExp result construction), and is critical for benchmark
performance.

The CL introducing ToDirectStringAssembler caused a couple of
regressions which this is intended to fix by adding a fast path for
sequential strings.

BUG=chromium:702246

Review-Url: https://codereview.chromium.org/2754933003
Cr-Commit-Position: refs/heads/master@{#43897}

src/code-stub-assembler.cc

@@ -3367,7 +3367,8 @@ Node* ToDirectStringAssembler::TryToDirect(Label* if_bailout) {
   Label if_isthin(this);
   Label out(this);
 
-  Goto(&dispatch);
+  Branch(IsSequentialStringInstanceType(var_instance_type_.value()), &out,
+         &dispatch);
 
   // Dispatch based on string representation.
   Bind(&dispatch);
@@ -3444,7 +3445,7 @@ Node* ToDirectStringAssembler::TryToSequential(StringPointerKind ptr_kind,
   CHECK(ptr_kind == PTR_TO_DATA || ptr_kind == PTR_TO_STRING);
 
   Variable var_result(this, MachineType::PointerRepresentation());
-  Label out(this), if_issequential(this), if_isexternal(this);
+  Label out(this), if_issequential(this), if_isexternal(this, Label::kDeferred);
   Branch(is_external(), &if_isexternal, &if_issequential);
 
   Bind(&if_issequential);