Fix bug in x86 memcpy / memset lowering. If there are trailing bytes not handled by rep instructions, a new memcpy / memset is introduced for them. However, since source / destination addresses are already adjusted, their offsets should be zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50239 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 20 insertions, 11 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 9ba1f53..a0fd3d9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -4822,7 +4822,7 @@ X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
                                       DAG.getConstant(Offset, AddrVT)),
                           Src,
                           DAG.getConstant(BytesLeft, SizeVT),
-                          Align, DstSV, Offset);
+                          Align, DstSV, 0);
   }
 
   // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
@@ -4847,8 +4847,6 @@ X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
   if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
     return SDOperand();
 
-  SmallVector<SDOperand, 4> Results;
-
   MVT::ValueType AVT;
   unsigned BytesLeft = 0;
   if (Align >= 8 && Subtarget->is64Bit())
@@ -4881,25 +4879,24 @@ X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
   Ops.push_back(Chain);
   Ops.push_back(DAG.getValueType(AVT));
   Ops.push_back(InFlag);
-  Results.push_back(DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size()));
+  SDOperand RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
 
+  SmallVector<SDOperand, 4> Results;
+  Results.push_back(RepMovs);
   if (BytesLeft) {
     // Handle the last 1 - 7 bytes.
     unsigned Offset = SizeVal - BytesLeft;
     MVT::ValueType DstVT = Dst.getValueType();
     MVT::ValueType SrcVT = Src.getValueType();
     MVT::ValueType SizeVT = Size.getValueType();
-
-    Results.push_back(DAG.getMemcpy(Chain, 
+    Results.push_back(DAG.getMemcpy(Chain,
                                     DAG.getNode(ISD::ADD, DstVT, Dst,
-                                                DAG.getConstant(Offset,
-                                                                DstVT)),
+                                                DAG.getConstant(Offset, DstVT)),
                                     DAG.getNode(ISD::ADD, SrcVT, Src,
-                                                DAG.getConstant(Offset,
-                                                                SrcVT)),
+                                                DAG.getConstant(Offset, SrcVT)),
                                     DAG.getConstant(BytesLeft, SizeVT),
                                     Align, AlwaysInline,
-                                    DstSV, Offset, SrcSV, Offset));
+                                    DstSV, 0, SrcSV, 0));
   }
 
   return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
diff --git a/test/CodeGen/X86/2008-04-24-MemCpyBug.ll b/test/CodeGen/X86/2008-04-24-MemCpyBug.ll
new file mode 100644
index 0000000..09fdc70
--- /dev/null
+++ b/test/CodeGen/X86/2008-04-24-MemCpyBug.ll
@@ -0,0 +1,12 @@
+; RUN: llvm-as < %s | llc -march=x86 | not grep 120
+; Don't accidentally add the offset twice for trailing bytes.
+
+	%struct.S63 = type { [63 x i8] }
+@g1s63 = external global %struct.S63		; <%struct.S63*> [#uses=1]
+
+declare void @test63(%struct.S63* byval align 4 ) nounwind 
+
+define void @testit63_entry_2E_ce() nounwind  {
+	tail call void @test63( %struct.S63* byval align 4  @g1s63 ) nounwind 
+	ret void
+}