[llvm] r341675 - [X86] Teach X86DAGToDAGISel::foldLoadStoreIntoMemOperand to handle loads in operand 1 of commutable operations.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Fri Sep  7 09:27:55 2018
New Revision: 341675

URL: http://llvm.org/viewvc/llvm-project?rev=341675&view=rev
Log:
[X86] Teach X86DAGToDAGISel::foldLoadStoreIntoMemOperand to handle loads in operand 1 of commutable operations.

Previously we only handled loads in operand 0, but nothing guarantees the load will be operand 0 for commutable operations.

Differential Revision: https://reviews.llvm.org/D51768

Modified:
    llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
    llvm/trunk/test/CodeGen/X86/addcarry.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp?rev=341675&r1=341674&r2=341675&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp Fri Sep  7 09:27:55 2018
@@ -2225,6 +2225,7 @@ static bool hasNoCarryFlagUses(SDNode *N
 /// the {load; op; store} to modify transformation.
 static bool isFusableLoadOpStorePattern(StoreSDNode *StoreNode,
                                         SDValue StoredVal, SelectionDAG *CurDAG,
+                                        unsigned LoadOpNo,
                                         LoadSDNode *&LoadNode,
                                         SDValue &InputChain) {
   // Is the stored value result 0 of the operation?
@@ -2237,7 +2238,7 @@ static bool isFusableLoadOpStorePattern(
   if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
     return false;
 
-  SDValue Load = StoredVal->getOperand(0);
+  SDValue Load = StoredVal->getOperand(LoadOpNo);
   // Is the stored value a non-extending and non-indexed load?
   if (!ISD::isNormalLoad(Load.getNode())) return false;
 
@@ -2366,26 +2367,39 @@ bool X86DAGToDAGISel::foldLoadStoreIntoM
   if (MemVT != MVT::i64 && MemVT != MVT::i32 && MemVT != MVT::i16 &&
       MemVT != MVT::i8)
     return false;
+
+  bool IsCommutable = false;
   switch (Opc) {
   default:
     return false;
   case X86ISD::INC:
   case X86ISD::DEC:
-  case X86ISD::ADD:
-  case X86ISD::ADC:
   case X86ISD::SUB:
   case X86ISD::SBB:
+    break;
+  case X86ISD::ADD:
+  case X86ISD::ADC:
   case X86ISD::AND:
   case X86ISD::OR:
   case X86ISD::XOR:
+    IsCommutable = true;
     break;
   }
 
+  unsigned LoadOpNo = 0;
   LoadSDNode *LoadNode = nullptr;
   SDValue InputChain;
-  if (!isFusableLoadOpStorePattern(StoreNode, StoredVal, CurDAG, LoadNode,
-                                   InputChain))
-    return false;
+  if (!isFusableLoadOpStorePattern(StoreNode, StoredVal, CurDAG, LoadOpNo,
+                                   LoadNode, InputChain)) {
+    if (!IsCommutable)
+      return false;
+
+    // This operation is commutable, try the other operand.
+    LoadOpNo = 1;
+    if (!isFusableLoadOpStorePattern(StoreNode, StoredVal, CurDAG, LoadOpNo,
+                                     LoadNode, InputChain))
+      return false;
+  }
 
   SDValue Base, Scale, Index, Disp, Segment;
   if (!selectAddr(LoadNode, LoadNode->getBasePtr(), Base, Scale, Index, Disp,
@@ -2503,7 +2517,7 @@ bool X86DAGToDAGISel::foldLoadStoreIntoM
     };
 
     unsigned NewOpc = SelectRegOpcode(Opc);
-    SDValue Operand = StoredVal->getOperand(1);
+    SDValue Operand = StoredVal->getOperand(1-LoadOpNo);
 
     // See if the operand is a constant that we can fold into an immediate
     // operand.

Modified: llvm/trunk/test/CodeGen/X86/addcarry.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/addcarry.ll?rev=341675&r1=341674&r2=341675&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/addcarry.ll (original)
+++ llvm/trunk/test/CodeGen/X86/addcarry.ll Fri Sep  7 09:27:55 2018
@@ -30,9 +30,8 @@ entry:
 define void @add128_rmw2(i128 %a, i128* %b) nounwind {
 ; CHECK-LABEL: add128_rmw2:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    addq (%rdx), %rdi
+; CHECK-NEXT:    addq %rdi, (%rdx)
 ; CHECK-NEXT:    adcq %rsi, 8(%rdx)
-; CHECK-NEXT:    movq %rdi, (%rdx)
 ; CHECK-NEXT:    retq
 entry:
   %0 = load i128, i128* %b
@@ -77,13 +76,10 @@ entry:
 define void @add256_rmw2(i256 %a, i256* %b) nounwind {
 ; CHECK-LABEL: add256_rmw2:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    addq (%r8), %rdi
-; CHECK-NEXT:    adcq 8(%r8), %rsi
-; CHECK-NEXT:    adcq 16(%r8), %rdx
+; CHECK-NEXT:    addq %rdi, (%r8)
+; CHECK-NEXT:    adcq %rsi, 8(%r8)
+; CHECK-NEXT:    adcq %rdx, 16(%r8)
 ; CHECK-NEXT:    adcq %rcx, 24(%r8)
-; CHECK-NEXT:    movq %rdi, (%r8)
-; CHECK-NEXT:    movq %rsi, 8(%r8)
-; CHECK-NEXT:    movq %rdx, 16(%r8)
 ; CHECK-NEXT:    retq
 entry:
   %0 = load i256, i256* %b