Merge pull request #4825 from bylaws/unityomg

Frontend: Force acq/rel semantics for known Unity ringbuffer offsets

FEXCore/Source/Interface/Core/Core.cpp

@@ -603,11 +603,17 @@ ContextImpl::GenerateIR(FEXCore::Core::InternalThreadState* Thread, uint64_t Gue
           auto Fn = TableInfo->OpcodeDispatcher.OpDispatch;
           Thread->OpDispatcher->ResetHandledLock();
           Thread->OpDispatcher->ResetDecodeFailure();
-          IR::ForceTSOMode ForceTSO =
-            BlockInForceTSOValidRange ?
-              (InstForceTSOIt != ForceTSOInstructions.end() && *InstForceTSOIt == InstAddress ? IR::ForceTSOMode::ForceEnabled :
-                                                                                                IR::ForceTSOMode::ForceDisabled) :
-              IR::ForceTSOMode::NoOverride;
+          IR::ForceTSOMode ForceTSO = IR::ForceTSOMode::NoOverride;
+          if (BlockInForceTSOValidRange) {
+            if (InstForceTSOIt != ForceTSOInstructions.end() && *InstForceTSOIt == InstAddress) {
+              ForceTSO = IR::ForceTSOMode::ForceEnabled;
+            } else {
+              ForceTSO = IR::ForceTSOMode::ForceDisabled;
+            }
+          } else if (DecodedInfo->ForceTSO) {
+            ForceTSO = IR::ForceTSOMode::ForceEnabled;
+          }
+
           Thread->OpDispatcher->SetForceTSO(ForceTSO);
           std::invoke(Fn, Thread->OpDispatcher, DecodedInfo);
           if (Thread->OpDispatcher->HadDecodeFailure()) {

FEXCore/Source/Interface/Core/Frontend.cpp

@@ -1046,6 +1046,22 @@ Decoder::DecodedBlockStatus Decoder::DecodeInstruction(uint64_t PC) {
     return DecodedBlockStatus::INVALID_INST;
   }
 
+  if (CTX->AreMonoHacksActive()) {
+    // Unity uses a standard SPSC ringbuffer with cached read/write pointers and thread waiting flags at the following
+    // offsets, which are consistent between 32-bit and 64-bit Unity versions from 2015 onwards.
+    auto IsKnownAtomicDisplacement = [](uint64_t Displacement) {
+      return Displacement == 0x80 || Displacement == 0x84 || Displacement == 0xC0 || Displacement == 0xC4;
+    };
+
+    if (DecodeInst->OP == 0x8b && DecodeInst->Src[0].IsGPRIndirect() &&
+        IsKnownAtomicDisplacement(DecodeInst->Src[0].Data.GPRIndirect.Displacement)) {
+      DecodeInst->ForceTSO = true;
+    }
+    if (DecodeInst->OP == 0x89 && DecodeInst->Dest.IsGPRIndirect() && IsKnownAtomicDisplacement(DecodeInst->Dest.Data.GPRIndirect.Displacement)) {
+      DecodeInst->ForceTSO = true;
+    }
+  }
+
   return DecodedBlockStatus::SUCCESS;
 }
 

FEXCore/Source/Interface/Core/X86Tables/X86Tables.h

@@ -200,6 +200,7 @@ struct DecodedInst {
   uint8_t LastEscapePrefix;
   bool DecodedModRM;
   bool DecodedSIB;
+  bool ForceTSO;
 };
 
 union ModRMDecoded {