Fix a serious X86 instruction selection bug. In

X86DAGToDAGISel::PreprocessISelDAG(), isel is moving load inside
callseq_start / callseq_end so it can be folded into a call. This can
create a cycle in the DAG when the call is glued to a copytoreg. We
have been lucky this hasn't caused too many issues because the pre-ra
scheduler has special handling of call sequences. However, it has
caused a crash in a specific tailcall case.

rdar://12393897

llvm-svn: 165072

diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
index b4c7154..98778c3 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -362,7 +362,7 @@ X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
 /// load's chain operand and move load below the call's chain operand.
 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
-                                  SDValue Call, SDValue OrigChain) {
+                               SDValue Call, SDValue OrigChain) {
   SmallVector<SDValue, 8> Ops;
   SDValue Chain = OrigChain.getOperand(0);
   if (Chain.getNode() == Load.getNode())
@@ -386,11 +386,22 @@ static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
   CurDAG->UpdateNodeOperands(OrigChain.getNode(), &Ops[0], Ops.size());
   CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
                              Load.getOperand(1), Load.getOperand(2));
+
+  bool IsGlued = Call.getOperand(0).getNode()->getGluedUser() == Call.getNode();
+  unsigned NumOps = Call.getNode()->getNumOperands();
   Ops.clear();
   Ops.push_back(SDValue(Load.getNode(), 1));
-  for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
+  for (unsigned i = 1, e = NumOps; i != e; ++i)
     Ops.push_back(Call.getOperand(i));
-  CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], Ops.size());
+  if (!IsGlued)
+    CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], NumOps);
+  else
+    // If call's chain was glued to the call (tailcall), and now the load
+    // is moved between them. Remove the glue to avoid a cycle (where the
+    // call is glued to its old chain and the load is using the old chain
+    // as its new chain).
+    CurDAG->MorphNodeTo(Call.getNode(), Call.getOpcode(),
+                        Call.getNode()->getVTList(), &Ops[0], NumOps-1);
 }
 
 /// isCalleeLoad - Return true if call address is a load and it can be

diff --git a/llvm/test/CodeGen/X86/2012-10-02-DAGCycle.ll b/llvm/test/CodeGen/X86/2012-10-02-DAGCycle.ll
new file mode 100644 (file)
index 0000000..9d2b7ea
--- /dev/null
+++ b/llvm/test/CodeGen/X86/2012-10-02-DAGCycle.ll
@@ -0,0 +1,16 @@
+; RUN: llc -mtriple=i386-apple-macosx -relocation-model=pic < %s
+; rdar://12393897
+
+%TRp = type { i32, %TRH*, i32, i32 }
+%TRH = type { i8*, i8*, i8*, i8*, {}* }
+
+define i32 @t(%TRp* inreg %rp) nounwind optsize ssp {
+entry:
+  %handler = getelementptr inbounds %TRp* %rp, i32 0, i32 1
+  %0 = load %TRH** %handler, align 4
+  %sync = getelementptr inbounds %TRH* %0, i32 0, i32 4
+  %sync12 = load {}** %sync, align 4
+  %1 = bitcast {}* %sync12 to i32 (%TRp*)*
+  %call = tail call i32 %1(%TRp* inreg %rp) nounwind optsize
+  ret i32 %call
+}