Intrinsify Interlocked.And and Interlocked.Or on XARCH (#96258)

Author: EgorBo

src/coreclr/jit/codegenxarch.cpp

@@ -2047,12 +2047,9 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
 
         case GT_XCHG:
         case GT_XADD:
-            genLockedInstructions(treeNode->AsOp());
-            break;
-
         case GT_XORR:
         case GT_XAND:
-            NYI("Interlocked.Or and Interlocked.And aren't implemented for x86 yet.");
+            genLockedInstructions(treeNode->AsOp());
             break;
 
         case GT_MEMORYBARRIER:
@@ -4413,7 +4410,7 @@ void CodeGen::genCodeForLockAdd(GenTreeOp* node)
 //
 void CodeGen::genLockedInstructions(GenTreeOp* node)
 {
-    assert(node->OperIs(GT_XADD, GT_XCHG));
+    assert(node->OperIs(GT_XADD, GT_XCHG, GT_XORR, GT_XAND));
 
     GenTree* addr = node->gtGetOp1();
     GenTree* data = node->gtGetOp2();
@@ -4425,6 +4422,56 @@ void CodeGen::genLockedInstructions(GenTreeOp* node)
 
     genConsumeOperands(node);
 
+    if (node->OperIs(GT_XORR, GT_XAND))
+    {
+        const instruction ins = node->OperIs(GT_XORR) ? INS_or : INS_and;
+
+        if (node->IsUnusedValue())
+        {
+            // If value is not used we can emit a short form:
+            //
+            //    lock
+            //    or/and  dword ptr [addrReg], val
+            //
+            instGen(INS_lock);
+            GetEmitter()->emitIns_AR_R(ins, size, data->GetRegNum(), addr->GetRegNum(), 0);
+        }
+        else
+        {
+            // When value is used (it's the original value of the memory location)
+            // we fallback to cmpxchg-loop idiom.
+
+            // for cmpxchg we need to keep the original value in RAX
+            assert(node->GetRegNum() == REG_RAX);
+
+            //    mov     RAX, dword ptr [addrReg]
+            //.LOOP:
+            //    mov     tmp, RAX
+            //    or/and  tmp, val
+            //    lock
+            //    cmpxchg dword ptr [addrReg], tmp
+            //    jne    .LOOP
+            //    ret
+
+            // Extend liveness of addr
+            gcInfo.gcMarkRegPtrVal(addr->GetRegNum(), addr->TypeGet());
+
+            const regNumber tmpReg = node->GetSingleTempReg();
+            GetEmitter()->emitIns_R_AR(INS_mov, size, REG_RAX, addr->GetRegNum(), 0);
+            BasicBlock* loop = genCreateTempLabel();
+            genDefineTempLabel(loop);
+            GetEmitter()->emitIns_Mov(INS_mov, size, tmpReg, REG_RAX, false);
+            GetEmitter()->emitIns_R_R(ins, size, tmpReg, data->GetRegNum());
+            instGen(INS_lock);
+            GetEmitter()->emitIns_AR_R(INS_cmpxchg, size, tmpReg, addr->GetRegNum(), 0);
+            inst_JMP(EJ_jne, loop);
+
+            gcInfo.gcMarkRegSetNpt(genRegMask(addr->GetRegNum()));
+            genProduceReg(node);
+        }
+        return;
+    }
+
     // If the destination register is different from the data register then we need
     // to first move the data to the target register. Make sure we don't overwrite
     // the address, the register allocator should have taken care of this.

src/coreclr/jit/importercalls.cpp

@@ -3231,14 +3231,20 @@ GenTree* Compiler::impIntrinsic(GenTree*                newobjThis,
                 break;
             }
 
-#if defined(TARGET_ARM64) || defined(TARGET_RISCV64)
-            // Intrinsify Interlocked.Or and Interlocked.And only for arm64-v8.1 (and newer) and for RV64A
-            // TODO-CQ: Implement for XArch (https://github.com/dotnet/runtime/issues/32239).
+#if defined(TARGET_ARM64) || defined(TARGET_RISCV64) || defined(TARGET_XARCH)
             case NI_System_Threading_Interlocked_Or:
             case NI_System_Threading_Interlocked_And:
             {
-                ARM64_ONLY(if (compOpportunisticallyDependsOn(InstructionSet_Atomics)))
+#if defined(TARGET_ARM64)
+                if (compOpportunisticallyDependsOn(InstructionSet_Atomics))
+#endif
                 {
+#if defined(TARGET_X86)
+                    if (genActualType(callType) == TYP_LONG)
+                    {
+                        break;
+                    }
+#endif
                     assert(sig->numArgs == 2);
                     GenTree*   op2 = impPopStack().val;
                     GenTree*   op1 = impPopStack().val;

src/coreclr/jit/lower.cpp

@@ -640,8 +640,6 @@ GenTree* Lowering::LowerNode(GenTree* node)
             CheckImmedAndMakeContained(node, node->AsOp()->gtOp2);
             break;
 #elif defined(TARGET_XARCH)
-        case GT_XORR:
-        case GT_XAND:
         case GT_XADD:
             if (node->IsUnusedValue())
             {

src/coreclr/jit/lsraxarch.cpp

@@ -436,6 +436,19 @@ int LinearScan::BuildNode(GenTree* tree)
 
         case GT_XORR:
         case GT_XAND:
+            if (!tree->IsUnusedValue())
+            {
+                // if tree's value is used, we'll emit a cmpxchg-loop idiom (requires RAX)
+                buildInternalIntRegisterDefForNode(tree, availableIntRegs & ~RBM_RAX);
+                BuildUse(tree->gtGetOp1(), availableIntRegs & ~RBM_RAX);
+                BuildUse(tree->gtGetOp2(), availableIntRegs & ~RBM_RAX);
+                BuildDef(tree, RBM_RAX);
+                buildInternalRegisterUses();
+                srcCount = 2;
+                assert(dstCount == 1);
+                break;
+            }
+            FALLTHROUGH;
         case GT_XADD:
         case GT_XCHG:
         {