[MERGE chakra-core#2307 @rajatd] Recover perf regression related to stack args

rajatd · rajatd · commit 582a9c3eb3b1 · 2017-01-05T11:47:47.000-08:00
Merge pull request chakra-core#2307 from rajatd:fixBox2dRegression A perf regression had crept in with a refactor I did a while ago. We dont have block-level merging of arguments object syms in the globopt. So, as much as possible, we try to keep the set of arguments object syms same across blocks. To this end, we ignore tracking the arguments object alias introduced by ByteCodeArgoutCapture. To do this, the globopt expects the argout instructions to be immediately preceding the InlineBuiltInStart instr for .apply. The refactor I did had broken this assumption. Reverting to that.
diff --git a/lib/Backend/Inline.cpp b/lib/Backend/Inline.cpp
@@ -2442,6 +2442,14 @@ IR::Instr * Inline::InlineApplyWithArgumentsObject(IR::Instr * callInstr, IR::In
     IR::Instr* builtInStartInstr;
     InsertInlineeBuiltInStartEndTags(callInstr, 3, &builtInStartInstr); //3 args (implicit this + explicit this + arguments = 3)
 
+    // Move argouts close to call. Globopt expects this for arguments object tracking.
+    IR::Instr* argInsertInstr = builtInStartInstr;
+    builtInStartInstr->IterateArgInstrs([&](IR::Instr* argInstr) {
+        argInstr->Move(argInsertInstr);
+        argInsertInstr = argInstr;
+        return false;
+    });
+
     IR::Instr *startCall = IR::Instr::New(Js::OpCode::StartCall, callInstr->m_func);
     startCall->SetDst(IR::RegOpnd::New(TyVar, callInstr->m_func));
     startCall->SetSrc1(IR::IntConstOpnd::New(2, TyInt32, callInstr->m_func)); //2 args (this pointer & ArgOut_A_From_StackArgs for this direct call to init
