Add xarch andn

src/coreclr/jit/emitxarch.cpp

@@ -423,6 +423,10 @@ bool emitter::AreFlagsSetToZeroCmp(regNumber reg, emitAttr opSize, genTreeOps tr
         case IF_RWR:
         case IF_RRD:
         case IF_RRW:
+        case IF_RWR_RRD_RRD:
+        case IF_RWR_RRD_MRD:
+        case IF_RWR_RRD_ARD:
+        case IF_RWR_RRD_SRD:
             break;
         default:
             return false;

src/coreclr/jit/instrsxarch.h

@@ -592,7 +592,7 @@ INST3(LAST_AVXVNNI_INSTRUCTION, "LAST_AVXVNNI_INSTRUCTION", IUM_WR, BAD_CODE, BA
 
 // BMI1
 INST3(FIRST_BMI_INSTRUCTION, "FIRST_BMI_INSTRUCTION", IUM_WR, BAD_CODE, BAD_CODE, BAD_CODE, INS_FLAGS_None)
-INST3(andn,             "andn",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF2),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Logical AND NOT
+INST3(andn,             "andn",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF2),                             Resets_OF      | Writes_SF     | Writes_ZF     | Undefined_AF  | Undefined_PF  | Resets_CF     | INS_Flags_IsDstDstSrcAVXInstruction)    // Logical AND NOT
 INST3(blsi,             "blsi",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF3),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Extract Lowest Set Isolated Bit
 INST3(blsmsk,           "blsmsk",           IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF3),                             INS_Flags_IsDstDstSrcAVXInstruction)    // Get Mask Up to Lowest Set Bit
 INST3(blsr,             "blsr",             IUM_WR, BAD_CODE,     BAD_CODE,     SSE38(0xF3),                             Resets_OF      | Writes_SF     | Writes_ZF     | Undefined_AF  | Undefined_PF  | Writes_CF     | INS_Flags_IsDstDstSrcAVXInstruction)    // Reset Lowest Set Bit

src/coreclr/jit/lower.cpp

@@ -139,7 +139,7 @@ GenTree* Lowering::LowerNode(GenTree* node)
         case GT_AND:
         case GT_OR:
         case GT_XOR:
-            return LowerBinaryArithmeticCommon(node->AsOp());
+            return LowerBinaryArithmetic(node->AsOp());
 
         case GT_MUL:
         case GT_MULHI:
@@ -5122,53 +5122,6 @@ GenTree* Lowering::LowerAdd(GenTreeOp* node)
     return nullptr;
 }
 
-//------------------------------------------------------------------------
-// LowerBinaryArithmeticCommon: lowers the given binary arithmetic node.
-//
-// Recognizes opportunities for using target-independent "combined" nodes
-// (currently AND_NOT on ARMArch). Performs containment checks.
-//
-// Arguments:
-//    node - the arithmetic node to lower
-//
-// Returns:
-//    The next node to lower.
-//
-GenTree* Lowering::LowerBinaryArithmeticCommon(GenTreeOp* binOp)
-{
-    // TODO-CQ-XArch: support BMI2 "andn" in codegen and condition
-    // this logic on the support for the instruction set on XArch.
-    CLANG_FORMAT_COMMENT_ANCHOR;
-
-#ifdef TARGET_ARMARCH
-    if (comp->opts.OptimizationEnabled() && binOp->OperIs(GT_AND))
-    {
-        GenTree* opNode  = nullptr;
-        GenTree* notNode = nullptr;
-        if (binOp->gtGetOp1()->OperIs(GT_NOT))
-        {
-            notNode = binOp->gtGetOp1();
-            opNode  = binOp->gtGetOp2();
-        }
-        else if (binOp->gtGetOp2()->OperIs(GT_NOT))
-        {
-            notNode = binOp->gtGetOp2();
-            opNode  = binOp->gtGetOp1();
-        }
-
-        if (notNode != nullptr)
-        {
-            binOp->gtOp1 = opNode;
-            binOp->gtOp2 = notNode->AsUnOp()->gtGetOp1();
-            binOp->ChangeOper(GT_AND_NOT);
-            BlockRange().Remove(notNode);
-        }
-    }
-#endif
-
-    return LowerBinaryArithmetic(binOp);
-}
-
 //------------------------------------------------------------------------
 // LowerUnsignedDivOrMod: Lowers a GT_UDIV/GT_UMOD node.
 //

src/coreclr/jit/lower.h

@@ -297,7 +297,6 @@ class Lowering final : public Phase
     void LowerStoreIndir(GenTreeStoreInd* node);
     GenTree* LowerAdd(GenTreeOp* node);
     GenTree* LowerMul(GenTreeOp* mul);
-    GenTree* LowerBinaryArithmeticCommon(GenTreeOp* binOp);
     GenTree* LowerBinaryArithmetic(GenTreeOp* binOp);
     bool LowerUnsignedDivOrMod(GenTreeOp* divMod);
     GenTree* LowerConstIntDivOrMod(GenTree* node);
@@ -344,7 +343,8 @@ class Lowering final : public Phase
     void LowerHWIntrinsicToScalar(GenTreeHWIntrinsic* node);
     void LowerHWIntrinsicGetElement(GenTreeHWIntrinsic* node);
     void LowerHWIntrinsicWithElement(GenTreeHWIntrinsic* node);
-    GenTree* TryLowerAndOpToResetLowestSetBit(GenTreeOp* binOp);
+    GenTree* TryLowerAndOpToResetLowestSetBit(GenTreeOp* andNode);
+    GenTree* TryLowerAndOpToAndNot(GenTreeOp* andNode);
 #elif defined(TARGET_ARM64)
     bool IsValidConstForMovImm(GenTreeHWIntrinsic* node);
     void LowerHWIntrinsicFusedMultiplyAddScalar(GenTreeHWIntrinsic* node);

src/coreclr/jit/lowerarmarch.cpp

@@ -292,6 +292,30 @@ GenTree* Lowering::LowerMul(GenTreeOp* mul)
 //
 GenTree* Lowering::LowerBinaryArithmetic(GenTreeOp* binOp)
 {
+    if (comp->opts.OptimizationEnabled() && binOp->OperIs(GT_AND))
+    {
+        GenTree* opNode  = nullptr;
+        GenTree* notNode = nullptr;
+        if (binOp->gtGetOp1()->OperIs(GT_NOT))
+        {
+            notNode = binOp->gtGetOp1();
+            opNode  = binOp->gtGetOp2();
+        }
+        else if (binOp->gtGetOp2()->OperIs(GT_NOT))
+        {
+            notNode = binOp->gtGetOp2();
+            opNode  = binOp->gtGetOp1();
+        }
+
+        if (notNode != nullptr)
+        {
+            binOp->gtOp1 = opNode;
+            binOp->gtOp2 = notNode->AsUnOp()->gtGetOp1();
+            binOp->ChangeOper(GT_AND_NOT);
+            BlockRange().Remove(notNode);
+        }
+    }
+
     ContainCheckBinary(binOp);
 
     return binOp->gtNext;

src/coreclr/jit/lowerxarch.cpp

@@ -162,6 +162,9 @@ GenTree* Lowering::LowerMul(GenTreeOp* mul)
 //------------------------------------------------------------------------
 // LowerBinaryArithmetic: lowers the given binary arithmetic node.
 //
+// Recognizes opportunities for using target-independent "combined" nodes
+// Performs containment checks.
+//
 // Arguments:
 //    node - the arithmetic node to lower
 //
@@ -173,10 +176,16 @@ GenTree* Lowering::LowerBinaryArithmetic(GenTreeOp* binOp)
 #ifdef FEATURE_HW_INTRINSICS
     if (comp->opts.OptimizationEnabled() && binOp->OperIs(GT_AND) && varTypeIsIntegral(binOp))
     {
-        GenTree* blsrNode = TryLowerAndOpToResetLowestSetBit(binOp);
-        if (blsrNode != nullptr)
+        GenTree* replacementNode = TryLowerAndOpToAndNot(binOp);
+        if (replacementNode != nullptr)
+        {
+            return replacementNode->gtNext;
+        }
+
+        replacementNode = TryLowerAndOpToResetLowestSetBit(binOp);
+        if (replacementNode != nullptr)
         {
-            return blsrNode->gtNext;
+            return replacementNode->gtNext;
         }
     }
 #endif
@@ -3726,14 +3735,16 @@ void Lowering::LowerHWIntrinsicToScalar(GenTreeHWIntrinsic* node)
 }
 
 //----------------------------------------------------------------------------------------------
-// Lowering::TryLowerAndOpToResetLowestSetBit: Lowers a tree AND(X, ADD(X, -1) to HWIntrinsic::ResetLowestSetBit
+// Lowering::TryLowerAndOpToResetLowestSetBit: Lowers a tree AND(X, ADD(X, -1)) to HWIntrinsic::ResetLowestSetBit
 //
 // Arguments:
 //    andNode - GT_AND node of integral type
 //
 // Return Value:
 //    Returns the replacement node if one is created else nullptr indicating no replacement
 //
+// Notes:
+//    Performs containment checks on the replacement node if one is created
 GenTree* Lowering::TryLowerAndOpToResetLowestSetBit(GenTreeOp* andNode)
 {
     assert(andNode->OperIs(GT_AND) && varTypeIsIntegral(andNode));
@@ -3802,6 +3813,86 @@ GenTree* Lowering::TryLowerAndOpToResetLowestSetBit(GenTreeOp* andNode)
     return blsrNode;
 }
 
+//----------------------------------------------------------------------------------------------
+// Lowering::TryLowerAndOpToAndNot: Lowers a tree AND(X, NOT(Y)) to HWIntrinsic::AndNot
+//
+// Arguments:
+//    andNode - GT_AND node of integral type
+//
+// Return Value:
+//    Returns the replacement node if one is created else nullptr indicating no replacement
+//
+// Notes:
+//    Performs containment checks on the replacement node if one is created
+GenTree* Lowering::TryLowerAndOpToAndNot(GenTreeOp* andNode)
+{
+    assert(andNode->OperIs(GT_AND) && varTypeIsIntegral(andNode));
+
+    GenTree* opNode  = nullptr;
+    GenTree* notNode = nullptr;
+    if (andNode->gtGetOp1()->OperIs(GT_NOT))
+    {
+        notNode = andNode->gtGetOp1();
+        opNode  = andNode->gtGetOp2();
+    }
+    else if (andNode->gtGetOp2()->OperIs(GT_NOT))
+    {
+        notNode = andNode->gtGetOp2();
+        opNode  = andNode->gtGetOp1();
+    }
+
+    if (opNode == nullptr)
+    {
+        return nullptr;
+    }
+
+    // We want to avoid using "andn" when one of the operands is both a source and the destination and is also coming
+    // from memory. In this scenario, we will get smaller and likely faster code by using the RMW encoding of `and`
+    if (IsBinOpInRMWStoreInd(andNode))
+    {
+        return nullptr;
+    }
+
+    NamedIntrinsic intrinsic;
+    if (andNode->TypeIs(TYP_LONG) && comp->compOpportunisticallyDependsOn(InstructionSet_BMI1_X64))
+    {
+        intrinsic = NamedIntrinsic::NI_BMI1_X64_AndNot;
+    }
+    else if (comp->compOpportunisticallyDependsOn(InstructionSet_BMI1))
+    {
+        intrinsic = NamedIntrinsic::NI_BMI1_AndNot;
+    }
+    else
+    {
+        return nullptr;
+    }
+
+    LIR::Use use;
+    if (!BlockRange().TryGetUse(andNode, &use))
+    {
+        return nullptr;
+    }
+
+    // note that parameter order for andn is ~y, x so these are purposefully reversed when creating the node
+    GenTreeHWIntrinsic* andnNode =
+        comp->gtNewScalarHWIntrinsicNode(andNode->TypeGet(), notNode->AsUnOp()->gtGetOp1(), opNode, intrinsic);
+
+    JITDUMP("Lower: optimize AND(X, NOT(Y)))\n");
+    DISPNODE(andNode);
+    JITDUMP("to:\n");
+    DISPNODE(andnNode);
+
+    use.ReplaceWith(andnNode);
+
+    BlockRange().InsertBefore(andNode, andnNode);
+    BlockRange().Remove(andNode);
+    BlockRange().Remove(notNode);
+
+    ContainCheckHWIntrinsic(andnNode);
+
+    return andnNode;
+}
+
 #endif // FEATURE_HW_INTRINSICS
 
 //----------------------------------------------------------------------------------------------