Revert last 3 commits

Author: abhishek-kaushik22

llvm/include/llvm/CodeGen/SDPatternMatch.h

@@ -537,12 +537,6 @@ m_Load(const T0_P &Ch, const T1_P &Ptr, const T2_P &Offset) {
       TernaryOpc_match<T0_P, T1_P, T2_P>(ISD::LOAD, Ch, Ptr, Offset));
 }
 
-template <typename T0_P, typename T1_P, typename T2_P>
-inline TernaryOpc_match<T0_P, T1_P, T2_P>
-m_Load(const T0_P &Ch, const T1_P &Ptr, const T2_P &Offset) {
-  return TernaryOpc_match<T0_P, T1_P, T2_P>(ISD::LOAD, Ch, Ptr, Offset);
-}
-
 template <typename T0_P, typename T1_P, typename T2_P>
 inline TernaryOpc_match<T0_P, T1_P, T2_P>
 m_InsertElt(const T0_P &Vec, const T1_P &Val, const T2_P &Idx) {

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -53392,34 +53392,38 @@ static SDValue combineMaskedStore(SDNode *N, SelectionDAG &DAG,
 static SDValue foldToMaskedStore(StoreSDNode *Store, SelectionDAG &DAG,
                                  const SDLoc &Dl,
                                  const X86Subtarget &Subtarget) {
-  using namespace llvm::SDPatternMatch;
-
   if (!Subtarget.hasAVX() && !Subtarget.hasAVX2() && !Subtarget.hasAVX512())
     return SDValue();
 
-  if (!Store->isSimple() || Store->isTruncatingStore())
+  if (!Store->isSimple())
     return SDValue();
 
   SDValue StoredVal = Store->getValue();
   SDValue StorePtr = Store->getBasePtr();
   SDValue StoreOffset = Store->getOffset();
-  EVT VT = Store->getMemoryVT();
+  EVT VT = StoredVal.getValueType();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   if (!TLI.isTypeLegal(VT) || !TLI.isOperationLegalOrCustom(ISD::MSTORE, VT))
     return SDValue();
 
-  SDValue Mask, TrueVec, LoadCh;
-  if (!sd_match(StoredVal,
-                m_VSelect(m_Value(Mask), m_Value(TrueVec),
-                          m_Load(m_Value(LoadCh), m_Specific(StorePtr),
-                                 m_Specific(StoreOffset)))))
+  if (StoredVal.getOpcode() != ISD::VSELECT)
     return SDValue();
 
-  LoadSDNode *Load = cast<LoadSDNode>(StoredVal.getOperand(2));
+  SDValue Mask = StoredVal.getOperand(0);
+  SDValue TrueVec = StoredVal.getOperand(1);
+  SDValue FalseVec = StoredVal.getOperand(2);
+
+  LoadSDNode *Load = cast<LoadSDNode>(FalseVec.getNode());
   if (!Load || !Load->isSimple())
     return SDValue();
 
+  SDValue LoadPtr = Load->getBasePtr();
+  SDValue LoadOffset = Load->getOffset();
+
+  if (StorePtr != LoadPtr || StoreOffset != LoadOffset)
+    return SDValue();
+
   auto IsSafeToFold = [](StoreSDNode *Store, LoadSDNode *Load) {
     std::queue<SDValue> Worklist;
 
@@ -53433,13 +53437,13 @@ static SDValue foldToMaskedStore(StoreSDNode *Store, SelectionDAG &DAG,
       if (!Node)
         return false;
 
-      if (Node == Load)
-        return true;
-
       if (const auto *MemNode = dyn_cast<MemSDNode>(Node))
         if (!MemNode->isSimple() || MemNode->writeMem())
           return false;
 
+      if (Node == Load)
+        return true;
+
       if (Node->getOpcode() == ISD::TokenFactor) {
         for (unsigned i = 0; i < Node->getNumOperands(); ++i)
           Worklist.push(Node->getOperand(i));
@@ -53455,8 +53459,8 @@ static SDValue foldToMaskedStore(StoreSDNode *Store, SelectionDAG &DAG,
     return SDValue();
 
   return DAG.getMaskedStore(Store->getChain(), Dl, TrueVec, StorePtr,
-                            StoreOffset, Mask, VT, Store->getMemOperand(),
-                            Store->getAddressingMode());
+                            StoreOffset, Mask, Store->getMemoryVT(),
+                            Store->getMemOperand(), Store->getAddressingMode());
 }
 
 static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
@@ -53723,9 +53727,6 @@ static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
                                    St->getMemOperand());
   }
 
-  if (SDValue MaskedStore = foldToMaskedStore(St, DAG, dl, Subtarget))
-    return MaskedStore;
-
   // Turn load->store of MMX types into GPR load/stores.  This avoids clobbering
   // the FP state in cases where an emms may be missing.
   // A preferable solution to the general problem is to figure out the right
@@ -53787,6 +53788,9 @@ static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
                         St->getMemOperand()->getFlags());
   }
 
+  if (SDValue MaskedStore = foldToMaskedStore(St, DAG, dl, Subtarget))
+    return MaskedStore;
+
   return SDValue();
 }
 

llvm/test/CodeGen/X86/combine-storetomstore.ll

@@ -218,7 +218,7 @@ define void @test_masked_store_intervening(<8 x i32> %x, ptr %ptr, <8 x i1> %cmp
 
 
 define void @test_masked_store_multiple(<8 x i32> %x, <8 x i32> %y, ptr %ptr1, ptr %ptr2, <8 x i1> %cmp, <8 x i1> %cmp2) {
-; AVX-LABEL: test_masked_store_multiple:
+; AVX-LABEL: foo:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpmovzxwd {{.*#+}} xmm4 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
 ; AVX-NEXT:    vpslld $31, %xmm4, %xmm4
@@ -237,7 +237,7 @@ define void @test_masked_store_multiple(<8 x i32> %x, <8 x i32> %y, ptr %ptr1, p
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
 ;
-; AVX2-LABEL: test_masked_store_multiple:
+; AVX2-LABEL: foo:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} ymm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
 ; AVX2-NEXT:    vpslld $31, %ymm2, %ymm2
@@ -250,7 +250,7 @@ define void @test_masked_store_multiple(<8 x i32> %x, <8 x i32> %y, ptr %ptr1, p
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
-; AVX512-LABEL: test_masked_store_multiple:
+; AVX512-LABEL: foo:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
 ; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0