diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp
index 394ad6c0109..74c11183e81 100644
--- a/lib/Analysis/ScalarEvolutionExpander.cpp
+++ b/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -49,6 +49,9 @@ Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
             // If the cast isn't the first instruction of the function, move it.
             if (BasicBlock::iterator(CI) != 
                 A->getParent()->getEntryBlock().begin()) {
+              // If the CastInst is the insert point, change the insert point.
+              if (CI == InsertPt) ++InsertPt;
+              // Splice the cast at the beginning of the entry block.
               CI->moveBefore(A->getParent()->getEntryBlock().begin());
             }
             return CI;
@@ -71,6 +74,8 @@ Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
             It = cast<InvokeInst>(I)->getNormalDest()->begin();
           while (isa<PHINode>(It)) ++It;
           if (It != BasicBlock::iterator(CI)) {
+            // If the CastInst is the insert point, change the insert point.
+            if (CI == InsertPt) ++InsertPt;
             // Splice the cast immediately after the operand in question.
             CI->moveBefore(It);
           }
diff --git a/test/Transforms/IndVarSimplify/casted-argument.ll b/test/Transforms/IndVarSimplify/casted-argument.ll
new file mode 100644
index 00000000000..ae41e3a8b7c
--- /dev/null
+++ b/test/Transforms/IndVarSimplify/casted-argument.ll
@@ -0,0 +1,24 @@
+; RUN: llvm-as < %s | opt -indvars -disable-output
+; PR4009
+
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
+target triple = "i386-pc-linux-gnu"
+
+define void @safe_bcopy(i8* %to) nounwind {
+entry:
+	%cmp11 = icmp ult i8* %to, null		; <i1> [#uses=1]
+	br i1 %cmp11, label %loop, label %return
+
+return:		; preds = %entry
+	ret void
+
+loop:		; preds = %loop, %if.else
+	%pn = phi i8* [ %ge, %loop ], [ null, %entry ]		; <i8*> [#uses=1]
+	%cp = ptrtoint i8* %to to i32		; <i32> [#uses=1]
+	%su = sub i32 0, %cp		; <i32> [#uses=1]
+	%ge = getelementptr i8* %pn, i32 %su		; <i8*> [#uses=2]
+	tail call void @bcopy(i8* %ge) nounwind
+	br label %loop
+}
+
+declare void @bcopy(i8* nocapture) nounwind