From 5eb0cecbc5dd370e33d4e0ab1abee7ce8597ca9c Mon Sep 17 00:00:00 2001
From: Dan Gohman <gohman@apple.com>
Date: Mon, 15 Sep 2008 19:46:03 +0000
Subject: [PATCH] Re-enable SelectionDAG CSE for calls. It matters in the case
 of libcalls, as in this testcase on ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56226 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/CodeGen/SelectionDAGNodes.h  |  3 ++-
 lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 26 +++++++++++++++-----
 test/CodeGen/ARM/cse-libcalls.ll          | 30 +++++++++++++++++++++++
 3 files changed, 52 insertions(+), 7 deletions(-)
 create mode 100644 test/CodeGen/ARM/cse-libcalls.ll

diff --git a/include/llvm/CodeGen/SelectionDAGNodes.h b/include/llvm/CodeGen/SelectionDAGNodes.h
index 85c464965dd..7cd9eef2abc 100644
--- a/include/llvm/CodeGen/SelectionDAGNodes.h
+++ b/include/llvm/CodeGen/SelectionDAGNodes.h
@@ -2205,7 +2205,8 @@ public:
 
   /// Set this call to not be marked as a tail call. Normally setter
   /// methods in SDNodes are unsafe because it breaks the CSE map,
-  /// but we don't CSE calls so it's ok in this case.
+  /// but we don't include the tail call flag for calls so it's ok
+  /// in this case.
   void setNotTailCall() { IsTailCall = false; }
 
   SDValue getChain() const { return getOperand(0); }
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 921d7b060b9..793f5c99972 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -423,6 +423,12 @@ static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
       ID.AddPointer(CP->getConstVal());
     break;
   }
+  case ISD::CALL: {
+    const CallSDNode *Call = cast<CallSDNode>(N);
+    ID.AddInteger(Call->getCallingConv());
+    ID.AddInteger(Call->isVarArg());
+    break;
+  }
   case ISD::LOAD: {
     const LoadSDNode *LD = cast<LoadSDNode>(N);
     ID.AddInteger(LD->getAddressingMode());
@@ -636,7 +642,6 @@ bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
   // not subject to CSE.
   if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
       !N->isMachineOpcode() &&
-      N->getOpcode() != ISD::CALL &&
       N->getOpcode() != ISD::DBG_LABEL &&
       N->getOpcode() != ISD::DBG_STOPPOINT &&
       N->getOpcode() != ISD::EH_LABEL &&
@@ -662,7 +667,6 @@ SDNode *SelectionDAG::AddNonLeafNodeToCSEMaps(SDNode *N) {
 
   switch (N->getOpcode()) {
   default: break;
-  case ISD::CALL:
   case ISD::HANDLENODE:
   case ISD::DBG_LABEL:
   case ISD::DBG_STOPPOINT:
@@ -3310,13 +3314,23 @@ SDValue
 SelectionDAG::getCall(unsigned CallingConv, bool IsVarArgs, bool IsTailCall,
                       SDVTList VTs,
                       const SDValue *Operands, unsigned NumOperands) {
-  // Do not CSE calls. Note that in addition to being a compile-time
-  // optimization (since attempting CSE of calls is unlikely to be
-  // meaningful), we actually depend on this behavior. CallSDNode can
-  // be mutated, which is only safe if calls are not CSE'd.
+  // Do not include isTailCall in the folding set profile.
+  FoldingSetNodeID ID;
+  AddNodeIDNode(ID, ISD::CALL, VTs, Operands, NumOperands);
+  ID.AddInteger(CallingConv);
+  ID.AddInteger(IsVarArgs);
+  void *IP = 0;
+  if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
+    // Instead of including isTailCall in the folding set, we just
+    // set the flag of the existing node.
+    if (!IsTailCall)
+      cast<CallSDNode>(E)->setNotTailCall();
+    return SDValue(E, 0);
+  }
   SDNode *N = NodeAllocator.Allocate<CallSDNode>();
   new (N) CallSDNode(CallingConv, IsVarArgs, IsTailCall,
                      VTs, Operands, NumOperands);
+  CSEMap.InsertNode(N, IP);
   AllNodes.push_back(N);
   return SDValue(N, 0);
 }
diff --git a/test/CodeGen/ARM/cse-libcalls.ll b/test/CodeGen/ARM/cse-libcalls.ll
new file mode 100644
index 00000000000..3b499a4021c
--- /dev/null
+++ b/test/CodeGen/ARM/cse-libcalls.ll
@@ -0,0 +1,30 @@
+; RUN: llvm-as < %s | llc -march=arm | grep {bl.\*__ltdf} | count 1
+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-f80:128:128"
+target triple = "i386-apple-darwin8"
+
+; Without CSE of libcalls, there are two calls in the output instead of one.
+
+define i32 @u_f_nonbon(double %lambda) nounwind {
+entry:
+	%tmp19.i.i = load double* null, align 4		; <double> [#uses=2]
+	%tmp6.i = fcmp olt double %tmp19.i.i, 1.000000e+00		; <i1> [#uses=1]
+	%dielectric.0.i = select i1 %tmp6.i, double 1.000000e+00, double %tmp19.i.i		; <double> [#uses=1]
+	%tmp10.i4 = fdiv double 0x4074C2D71F36262D, %dielectric.0.i		; <double> [#uses=1]
+	br i1 false, label %bb28.i, label %bb508.i
+
+bb28.i:		; preds = %bb28.i, %entry
+	br i1 false, label %bb502.loopexit.i, label %bb28.i
+
+bb.nph53.i:		; preds = %bb502.loopexit.i
+	%tmp354.i = sub double -0.000000e+00, %tmp10.i4		; <double> [#uses=0]
+	br label %bb244.i
+
+bb244.i:		; preds = %bb244.i, %bb.nph53.i
+	br label %bb244.i
+
+bb502.loopexit.i:		; preds = %bb28.i
+	br i1 false, label %bb.nph53.i, label %bb508.i
+
+bb508.i:		; preds = %bb502.loopexit.i, %entry
+	ret i32 1
+}