Propagate DAG node ordering during type legalization and instruction selection

A node's ordering is only propagated during legalization if (a) the new node does
not have an ordering (is not a CSE'd node), or (b) the new node has an ordering
that is higher than the node being legalized.

llvm-svn: 177465

llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp

@@ -619,6 +619,17 @@ void DAGTypeLegalizer::RemapValue(SDValue &N) {
   }
 }
 
+/// PropagateOrdering - Propagate SDNode ordering information from \p Old to
+/// \p New. Generally, this just means copying the ordering value, but if the
+/// new node is actually a recycled node with a lower ordering already, then
+/// we do not want to propagate the new (higher) ordering.
+void DAGTypeLegalizer::PropagateOrdering(SDNode *Old, SDNode *New) {
+  unsigned OldOrder = DAG.GetOrdering(Old);
+  unsigned NewOrder = DAG.GetOrdering(New);
+  if (NewOrder == 0 || (NewOrder > 0 && OldOrder < NewOrder))
+    DAG.AssignOrdering(New, OldOrder);
+}
+
 namespace {
   /// NodeUpdateListener - This class is a DAGUpdateListener that listens for
   /// updates to nodes and recomputes their ready state.
@@ -735,6 +746,9 @@ void DAGTypeLegalizer::SetPromotedInteger(SDValue Op, SDValue Result) {
   SDValue &OpEntry = PromotedIntegers[Op];
   assert(OpEntry.getNode() == 0 && "Node is already promoted!");
   OpEntry = Result;
+
+  // Propagate node ordering
+  PropagateOrdering(Op.getNode(), Result.getNode());
 }
 
 void DAGTypeLegalizer::SetSoftenedFloat(SDValue Op, SDValue Result) {
@@ -746,6 +760,9 @@ void DAGTypeLegalizer::SetSoftenedFloat(SDValue Op, SDValue Result) {
   SDValue &OpEntry = SoftenedFloats[Op];
   assert(OpEntry.getNode() == 0 && "Node is already converted to integer!");
   OpEntry = Result;
+
+  // Propagate node ordering
+  PropagateOrdering(Op.getNode(), Result.getNode());
 }
 
 void DAGTypeLegalizer::SetScalarizedVector(SDValue Op, SDValue Result) {
@@ -760,6 +777,9 @@ void DAGTypeLegalizer::SetScalarizedVector(SDValue Op, SDValue Result) {
   SDValue &OpEntry = ScalarizedVectors[Op];
   assert(OpEntry.getNode() == 0 && "Node is already scalarized!");
   OpEntry = Result;
+
+  // Propagate node ordering
+  PropagateOrdering(Op.getNode(), Result.getNode());
 }
 
 void DAGTypeLegalizer::GetExpandedInteger(SDValue Op, SDValue &Lo,
@@ -787,6 +807,10 @@ void DAGTypeLegalizer::SetExpandedInteger(SDValue Op, SDValue Lo,
   assert(Entry.first.getNode() == 0 && "Node already expanded");
   Entry.first = Lo;
   Entry.second = Hi;
+
+  // Propagate ordering
+  PropagateOrdering(Op.getNode(), Lo.getNode());
+  PropagateOrdering(Op.getNode(), Hi.getNode());
 }
 
 void DAGTypeLegalizer::GetExpandedFloat(SDValue Op, SDValue &Lo,
@@ -814,6 +838,10 @@ void DAGTypeLegalizer::SetExpandedFloat(SDValue Op, SDValue Lo,
   assert(Entry.first.getNode() == 0 && "Node already expanded");
   Entry.first = Lo;
   Entry.second = Hi;
+
+  // Propagate ordering
+  PropagateOrdering(Op.getNode(), Lo.getNode());
+  PropagateOrdering(Op.getNode(), Hi.getNode());
 }
 
 void DAGTypeLegalizer::GetSplitVector(SDValue Op, SDValue &Lo,
@@ -843,6 +871,10 @@ void DAGTypeLegalizer::SetSplitVector(SDValue Op, SDValue Lo,
   assert(Entry.first.getNode() == 0 && "Node already split");
   Entry.first = Lo;
   Entry.second = Hi;
+
+  // Propagate ordering
+  PropagateOrdering(Op.getNode(), Lo.getNode());
+  PropagateOrdering(Op.getNode(), Hi.getNode());
 }
 
 void DAGTypeLegalizer::SetWidenedVector(SDValue Op, SDValue Result) {
@@ -854,6 +886,9 @@ void DAGTypeLegalizer::SetWidenedVector(SDValue Op, SDValue Result) {
   SDValue &OpEntry = WidenedVectors[Op];
   assert(OpEntry.getNode() == 0 && "Node already widened!");
   OpEntry = Result;
+
+  // Propagate node ordering
+  PropagateOrdering(Op.getNode(), Result.getNode());
 }
 
 
@@ -919,8 +954,11 @@ bool DAGTypeLegalizer::CustomLowerNode(SDNode *N, EVT VT, bool LegalizeResult) {
   // Make everything that once used N's values now use those in Results instead.
   assert(Results.size() == N->getNumValues() &&
          "Custom lowering returned the wrong number of results!");
-  for (unsigned i = 0, e = Results.size(); i != e; ++i)
+  for (unsigned i = 0, e = Results.size(); i != e; ++i) {
     ReplaceValueWith(SDValue(N, i), Results[i]);
+    // Propagate node ordering
+    DAG.AssignOrdering(Results[i].getNode(), DAG.GetOrdering(N));
+  }
   return true;
 }
 

llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h

@@ -143,6 +143,7 @@ private:
   void ExpungeNode(SDNode *N);
   void PerformExpensiveChecks();
   void RemapValue(SDValue &N);
+  void PropagateOrdering(SDNode *Old, SDNode *New);
 
   // Common routines.
   SDValue BitConvertToInteger(SDValue Op);

llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -785,8 +785,10 @@ void SelectionDAGISel::DoInstructionSelection() {
       if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE)
         continue;
       // Replace node.
-      if (ResNode)
+      if (ResNode) {
+        CurDAG->AssignOrdering(ResNode, CurDAG->GetOrdering(Node));
         ReplaceUses(Node, ResNode);
+      }
 
       // If after the replacement this node is not used any more,
       // remove this dead node.

llvm/test/CodeGen/NVPTX/sched1.ll

@@ -0,0 +1,31 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+; Ensure source scheduling is working
+
+define void @foo(i32* %a) {
+; CHECK: .func foo
+; CHECK: ld.u32
+; CHECK-NEXT: ld.u32
+; CHECK-NEXT: ld.u32
+; CHECK-NEXT: ld.u32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+  %ptr0 = getelementptr i32* %a, i32 0
+  %val0 = load i32* %ptr0
+  %ptr1 = getelementptr i32* %a, i32 1
+  %val1 = load i32* %ptr1
+  %ptr2 = getelementptr i32* %a, i32 2
+  %val2 = load i32* %ptr2
+  %ptr3 = getelementptr i32* %a, i32 3
+  %val3 = load i32* %ptr3
+
+  %t0 = add i32 %val0, %val1
+  %t1 = add i32 %t0, %val2
+  %t2 = add i32 %t1, %val3
+
+  store i32 %t2, i32* %a
+
+  ret void
+}
+

llvm/test/CodeGen/NVPTX/sched2.ll

@@ -0,0 +1,32 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+define void @foo(<2 x i32>* %a) {
+; CHECK: .func foo
+; CHECK: ld.v2.u32
+; CHECK-NEXT: ld.v2.u32
+; CHECK-NEXT: ld.v2.u32
+; CHECK-NEXT: ld.v2.u32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+; CHECK-NEXT: add.s32
+  %ptr0 = getelementptr <2 x i32>* %a, i32 0
+  %val0 = load <2 x i32>* %ptr0
+  %ptr1 = getelementptr <2 x i32>* %a, i32 1
+  %val1 = load <2 x i32>* %ptr1
+  %ptr2 = getelementptr <2 x i32>* %a, i32 2
+  %val2 = load <2 x i32>* %ptr2
+  %ptr3 = getelementptr <2 x i32>* %a, i32 3
+  %val3 = load <2 x i32>* %ptr3
+
+  %t0 = add <2 x i32> %val0, %val1
+  %t1 = add <2 x i32> %t0, %val2
+  %t2 = add <2 x i32> %t1, %val3
+
+  store <2 x i32> %t2, <2 x i32>* %a
+
+  ret void
+}
+

llvm/test/CodeGen/NVPTX/vector-loads.ll

@@ -9,7 +9,7 @@
 
 define void @foo(<2 x float>* %a) {
 ; CHECK: .func foo
-; CHECK: ld.v2.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v2.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}}
   %t1 = load <2 x float>* %a
   %t2 = fmul <2 x float> %t1, %t1
   store <2 x float> %t2, <2 x float>* %a
@@ -18,7 +18,7 @@ define void @foo(<2 x float>* %a) {
 
 define void @foo2(<4 x float>* %a) {
 ; CHECK: .func foo2
-; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}
   %t1 = load <4 x float>* %a
   %t2 = fmul <4 x float> %t1, %t1
   store <4 x float> %t2, <4 x float>* %a
@@ -27,8 +27,8 @@ define void @foo2(<4 x float>* %a) {
 
 define void @foo3(<8 x float>* %a) {
 ; CHECK: .func foo3
-; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}
-; CHECK-NEXT: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [%r{{[0-9]+}}+16];
+; CHECK-NEXT: ld.v4.f32 {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}
   %t1 = load <8 x float>* %a
   %t2 = fmul <8 x float> %t1, %t1
   store <8 x float> %t2, <8 x float>* %a
@@ -39,7 +39,7 @@ define void @foo3(<8 x float>* %a) {
 
 define void @foo4(<2 x i32>* %a) {
 ; CHECK: .func foo4
-; CHECK: ld.v2.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v2.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}}
   %t1 = load <2 x i32>* %a
   %t2 = mul <2 x i32> %t1, %t1
   store <2 x i32> %t2, <2 x i32>* %a
@@ -48,7 +48,7 @@ define void @foo4(<2 x i32>* %a) {
 
 define void @foo5(<4 x i32>* %a) {
 ; CHECK: .func foo5
-; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}
   %t1 = load <4 x i32>* %a
   %t2 = mul <4 x i32> %t1, %t1
   store <4 x i32> %t2, <4 x i32>* %a
@@ -57,8 +57,8 @@ define void @foo5(<4 x i32>* %a) {
 
 define void @foo6(<8 x i32>* %a) {
 ; CHECK: .func foo6
-; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}];
+; CHECK: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}
-; CHECK-NEXT: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}, [%r{{[0-9]+}}+16];
+; CHECK-NEXT: ld.v4.u32 {%r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}}
   %t1 = load <8 x i32>* %a
   %t2 = mul <8 x i32> %t1, %t1
   store <8 x i32> %t2, <8 x i32>* %a

llvm/test/CodeGen/X86/win_ftol2.ll

@@ -63,9 +63,9 @@ define i64 @double_ui64_2(double %x, double %y, double %z) nounwind {
 
   %1 = fdiv double %x, %y
   %2 = fsub double %x, %z
-  %3 = fptoui double %1 to i64
+  %3 = fptoui double %2 to i64
-  %4 = fptoui double %2 to i64
+  %4 = fptoui double %1 to i64
-  %5 = sub i64 %3, %4
+  %5 = sub i64 %4, %3
   ret i64 %5
 }
 
@@ -121,9 +121,9 @@ define {double, i64} @double_ui64_4(double %x, double %y) nounwind {
 ; FTOL_2: calll __ftol2
 ;; stack is %x
 
-  %1 = fptoui double %x to i64
+  %1 = fptoui double %y to i64
-  %2 = fptoui double %y to i64
+  %2 = fptoui double %x to i64
-  %3 = sub i64 %1, %2
+  %3 = sub i64 %2, %1
   %4 = insertvalue {double, i64} undef, double %x, 0
   %5 = insertvalue {double, i64} %4, i64 %3, 1
   ret {double, i64} %5