Merge pull request #3894 from hthh/optimize-fifo-bug

Jit: Check the FIFO on EIEIO instructions

Source/Core/Core/PowerPC/Jit64/Jit.cpp

@@ -598,6 +598,7 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBloc
   js.isLastInstruction = false;
   js.blockStart = em_address;
   js.fifoBytesThisBlock = 0;
+  js.mustCheckFifo = false;
   js.curBlock = b;
   js.numLoadStoreInst = 0;
   js.numFloatingPointInst = 0;
@@ -723,9 +724,11 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBloc
         js.fifoWriteAddresses.find(ops[i].address) != js.fifoWriteAddresses.end();
 
     // Gather pipe writes using an immediate address are explicitly tracked.
-    if (jo.optimizeGatherPipe && js.fifoBytesThisBlock >= 32)
+    if (jo.optimizeGatherPipe && (js.fifoBytesThisBlock >= 32 || js.mustCheckFifo))
     {
-      js.fifoBytesThisBlock -= 32;
+      if (js.fifoBytesThisBlock >= 32)
+        js.fifoBytesThisBlock -= 32;
+      js.mustCheckFifo = false;
       BitSet32 registersInUse = CallerSavedRegistersInUse();
       ABI_PushRegistersAndAdjustStack(registersInUse, 0);
       ABI_CallFunction((void*)&GPFifo::FastCheckGatherPipe);

Source/Core/Core/PowerPC/Jit64/Jit.h

@@ -241,4 +241,6 @@ public:
   void stmw(UGeckoInstruction inst);
 
   void dcbx(UGeckoInstruction inst);
+
+  void eieio(UGeckoInstruction inst);
 };

Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp

@@ -309,7 +309,7 @@ static GekkoOPTemplate table31[] = {
     // Unused instructions on GC
     {310, &Jit64::FallBackToInterpreter},  // eciwx
     {438, &Jit64::FallBackToInterpreter},  // ecowx
-    {854, &Jit64::DoNothing},              // eieio
+    {854, &Jit64::eieio},                  // eieio
     {306, &Jit64::FallBackToInterpreter},  // tlbie
     {566, &Jit64::DoNothing},              // tlbsync
 };

Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -598,3 +598,15 @@ void Jit64::stmw(UGeckoInstruction inst)
   }
   gpr.UnlockAllX();
 }
+
+void Jit64::eieio(UGeckoInstruction inst)
+{
+  INSTRUCTION_START
+  JITDISABLE(bJITLoadStoreOff);
+
+  // optimizeGatherPipe generally postpones FIFO checks to the end of the JIT block,
+  // which is generally safe. However postponing FIFO writes across eieio instructions
+  // is incorrect (would crash NBA2K11 strap screen if we improve our FIFO detection).
+  if (jo.optimizeGatherPipe && js.fifoBytesThisBlock > 0)
+    js.mustCheckFifo = true;
+}

Source/Core/Core/PowerPC/JitArm64/Jit.cpp

@@ -405,6 +405,7 @@ const u8* JitArm64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitB
   js.assumeNoPairedQuantize = false;
   js.blockStart = em_address;
   js.fifoBytesThisBlock = 0;
+  js.mustCheckFifo = false;
   js.downcountAmount = 0;
   js.skipInstructions = 0;
   js.curBlock = b;
@@ -491,9 +492,11 @@ const u8* JitArm64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitB
     bool gatherPipeIntCheck =
         jit->js.fifoWriteAddresses.find(ops[i].address) != jit->js.fifoWriteAddresses.end();
 
-    if (jo.optimizeGatherPipe && js.fifoBytesThisBlock >= 32)
+    if (jo.optimizeGatherPipe && (js.fifoBytesThisBlock >= 32 || js.mustCheckFifo))
     {
-      js.fifoBytesThisBlock -= 32;
+      if (js.fifoBytesThisBlock >= 32)
+        js.fifoBytesThisBlock -= 32;
+      js.mustCheckFifo = false;
 
       gpr.Lock(W30);
       BitSet32 regs_in_use = gpr.GetCallerSavedUsed();

Source/Core/Core/PowerPC/JitArm64/Jit.h

@@ -116,6 +116,7 @@ public:
   void dcbx(UGeckoInstruction inst);
   void dcbt(UGeckoInstruction inst);
   void dcbz(UGeckoInstruction inst);
+  void eieio(UGeckoInstruction inst);
 
   // LoadStore floating point
   void lfXX(UGeckoInstruction inst);

Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp

@@ -853,3 +853,15 @@ void JitArm64::dcbz(UGeckoInstruction inst)
 
   gpr.Unlock(W0);
 }
+
+void JitArm64::eieio(UGeckoInstruction inst)
+{
+  INSTRUCTION_START
+  JITDISABLE(bJITLoadStoreOff);
+
+  // optimizeGatherPipe generally postpones FIFO checks to the end of the JIT block,
+  // which is generally safe. However postponing FIFO writes across eieio instructions
+  // is incorrect (would crash NBA2K11 strap screen if we improve our FIFO detection).
+  if (jo.optimizeGatherPipe && js.fifoBytesThisBlock > 0)
+    js.mustCheckFifo = true;
+}

Source/Core/Core/PowerPC/JitArm64/JitArm64_Tables.cpp

@@ -314,7 +314,7 @@ static GekkoOPTemplate table31[] = {
     // Unused instructions on GC
     {310, &JitArm64::FallBackToInterpreter},  // eciwx
     {438, &JitArm64::FallBackToInterpreter},  // ecowx
-    {854, &JitArm64::DoNothing},              // eieio
+    {854, &JitArm64::eieio},                  // eieio
     {306, &JitArm64::FallBackToInterpreter},  // tlbie
     {566, &JitArm64::DoNothing},              // tlbsync
 };

Source/Core/Core/PowerPC/JitCommon/JitBase.h

@@ -103,6 +103,7 @@ protected:
     bool generatingTrampoline = false;
     u8* trampolineExceptionHandler;
 
+    bool mustCheckFifo;
     int fifoBytesThisBlock;
 
     PPCAnalyst::BlockStats st;