Reduce impact of HParameter inputs on HPhi representation selection

This is a follow-up to r25153.

BUG=v8:3766
LOG=n

Review URL: https://codereview.chromium.org/807273003

Cr-Commit-Position: refs/heads/master@{#25885}

src/hydrogen-instructions.cc

@@ -4501,18 +4501,24 @@ void HPhi::SimplifyConstantInputs() {
 
 void HPhi::InferRepresentation(HInferRepresentationPhase* h_infer) {
   DCHECK(CheckFlag(kFlexibleRepresentation));
-  Representation new_rep = RepresentationFromInputs();
-  UpdateRepresentation(new_rep, h_infer, "inputs");
-  new_rep = RepresentationFromUses();
+  Representation new_rep = RepresentationFromUses();
   UpdateRepresentation(new_rep, h_infer, "uses");
+  new_rep = RepresentationFromInputs();
+  UpdateRepresentation(new_rep, h_infer, "inputs");
   new_rep = RepresentationFromUseRequirements();
   UpdateRepresentation(new_rep, h_infer, "use requirements");
 }
 
 
 Representation HPhi::RepresentationFromInputs() {
-  Representation r = Representation::None();
+  bool has_type_feedback =
+      smi_non_phi_uses() + int32_non_phi_uses() + double_non_phi_uses() > 0;
+  Representation r = representation();
   for (int i = 0; i < OperandCount(); ++i) {
+    // Ignore conservative Tagged assumption of parameters if we have
+    // reason to believe that it's too conservative.
+    if (has_type_feedback && OperandAt(i)->IsParameter()) continue;
+
     r = r.generalize(OperandAt(i)->KnownOptimalRepresentation());
   }
   return r;

src/hydrogen-instructions.h

@@ -1641,6 +1641,11 @@ class HForceRepresentation FINAL : public HTemplateInstruction<1> {
 
   HValue* value() const { return OperandAt(0); }
 
+  Representation observed_input_representation(int index) OVERRIDE {
+    // We haven't actually *observed* this, but it's closer to the truth
+    // than 'None'.
+    return representation();  // Same as the output representation.
+  }
   Representation RequiredInputRepresentation(int index) OVERRIDE {
     return representation();  // Same as the output representation.
   }