JitArm64: Rewrite Exit functions.

The gpr must not be touched in the Exit functions as they are maybe conditional.
So just allocate everything here manually.

2 files changed, 47 insertions, 70 deletions

diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
index 7cc2c476c2..4b27c13f00 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
@@ -161,6 +161,7 @@ void JitArm64::FallBackToInterpreter(UGeckoInstruction inst)
       ARM64Reg WA = gpr.GetReg();
       LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(npc));
       WriteExceptionExit(WA);
+      gpr.Unlock(WA);
     }
     else
     {
@@ -174,6 +175,7 @@ void JitArm64::FallBackToInterpreter(UGeckoInstruction inst)
       FixupBranch c = B(CC_EQ);
       WriteExceptionExit(WA);
       SetJumpTarget(c);
+      gpr.Unlock(WA);
     }
   }
 
@@ -211,6 +213,7 @@ void JitArm64::HLEFunction(UGeckoInstruction inst)
   ARM64Reg WA = gpr.GetReg();
   LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(npc));
   WriteExit(WA);
+  gpr.Unlock(WA);
 }
 
 void JitArm64::DoNothing(UGeckoInstruction inst)
@@ -228,21 +231,16 @@ void JitArm64::Cleanup()
 {
   if (jo.optimizeGatherPipe && js.fifoBytesSinceCheck > 0)
   {
-    gpr.Lock(W0);
     MOVP2R(X0, &GPFifo::FastCheckGatherPipe);
     BLR(X0);
-    gpr.Unlock(W0);
   }
 }
 
 void JitArm64::DoDownCount()
 {
-  ARM64Reg WA = gpr.GetReg();
-  LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(downcount));
-  ARM64Reg WB = gpr.GetReg();
-  SUBSI2R(WA, WA, js.downcountAmount, WB);
-  STR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(downcount));
-  gpr.Unlock(WA, WB);
+  LDR(INDEX_UNSIGNED, W0, PPC_REG, PPCSTATE_OFF(downcount));
+  SUBSI2R(W0, W0, js.downcountAmount, W1);
+  STR(INDEX_UNSIGNED, W0, PPC_REG, PPCSTATE_OFF(downcount));
 }
 
 void JitArm64::ResetStack()
@@ -292,9 +290,7 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
 {
   Cleanup();
   DoDownCount();
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   LK &= m_enable_blr_optimization;
 
@@ -331,18 +327,15 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
 
 void JitArm64::WriteExit(Arm64Gen::ARM64Reg dest, bool LK, u32 exit_address_after_return)
 {
+  if (dest != DISPATCHER_PC)
+    MOV(DISPATCHER_PC, dest);
+
   Cleanup();
   DoDownCount();
+  EndTimeProfile(js.curBlock);
 
   LK &= m_enable_blr_optimization;
 
-  if (dest != DISPATCHER_PC)
-    MOV(DISPATCHER_PC, dest);
-  gpr.Unlock(dest);
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
-
   if (!LK)
   {
     B(dispatcher);
@@ -407,35 +400,28 @@ void JitArm64::WriteBLRExit(Arm64Gen::ARM64Reg dest)
     return;
   }
 
-  Cleanup();
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  if (dest != DISPATCHER_PC)
+    MOV(DISPATCHER_PC, dest);
 
-  ARM64Reg code = gpr.GetReg();
-  ARM64Reg pc = gpr.GetReg();
+  Cleanup();
+  EndTimeProfile(js.curBlock);
 
   // Check if {ARM_PC, PPC_PC} matches the current state.
-  LDP(INDEX_POST, EncodeRegTo64(code), EncodeRegTo64(pc), SP, 16);
-  CMP(pc, dest);
+  LDP(INDEX_POST, X2, X1, SP, 16);
+  CMP(W1, DISPATCHER_PC);
   FixupBranch no_match = B(CC_NEQ);
 
-  DoDownCount();
+  DoDownCount();  // overwrites X0 + X1
 
-  RET(EncodeRegTo64(code));
+  RET(X2);
 
   SetJumpTarget(no_match);
 
   DoDownCount();
 
-  if (dest != DISPATCHER_PC)
-    MOV(DISPATCHER_PC, dest);
-
   ResetStack();
 
   B(dispatcher);
-
-  gpr.Unlock(dest, pc, code);
 }
 
 void JitArm64::WriteExceptionExit(u32 destination, bool only_external)
@@ -458,39 +444,34 @@ void JitArm64::WriteExceptionExit(u32 destination, bool only_external)
 
   SetJumpTarget(no_exceptions);
 
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   B(dispatcher);
 }
 
 void JitArm64::WriteExceptionExit(ARM64Reg dest, bool only_external)
 {
+  if (dest != DISPATCHER_PC)
+    MOV(DISPATCHER_PC, dest);
+
   Cleanup();
   DoDownCount();
 
-  ARM64Reg WA = gpr.GetReg();
-  LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(Exceptions));
-  FixupBranch no_exceptions = CBZ(WA);
-  gpr.Unlock(WA);
+  LDR(INDEX_UNSIGNED, W30, PPC_REG, PPCSTATE_OFF(Exceptions));
+  FixupBranch no_exceptions = CBZ(W30);
 
-  STR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(pc));
-  STR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(npc));
+  STR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(pc));
+  STR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(npc));
   if (only_external)
-    MOVP2R(EncodeRegTo64(dest), &PowerPC::CheckExternalExceptions);
+    MOVP2R(EncodeRegTo64(DISPATCHER_PC), &PowerPC::CheckExternalExceptions);
   else
-    MOVP2R(EncodeRegTo64(dest), &PowerPC::CheckExceptions);
-  BLR(EncodeRegTo64(dest));
-  LDR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(npc));
+    MOVP2R(EncodeRegTo64(DISPATCHER_PC), &PowerPC::CheckExceptions);
+  BLR(EncodeRegTo64(DISPATCHER_PC));
+  LDR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(npc));
 
   SetJumpTarget(no_exceptions);
 
-  if (dest != DISPATCHER_PC)
-    MOV(DISPATCHER_PC, dest);
-  gpr.Unlock(dest);
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   B(dispatcher);
 }
@@ -518,32 +499,24 @@ void JitArm64::BeginTimeProfile(JitBlock* b)
 
 void JitArm64::EndTimeProfile(JitBlock* b)
 {
-  ARM64Reg WA = gpr.GetReg();
-  ARM64Reg XA = EncodeRegTo64(WA);
-  ARM64Reg WB = gpr.GetReg();
-  ARM64Reg XB = EncodeRegTo64(WB);
-  ARM64Reg WC = gpr.GetReg();
-  ARM64Reg XC = EncodeRegTo64(WC);
-  ARM64Reg WD = gpr.GetReg();
-  ARM64Reg XD = EncodeRegTo64(WD);
+  if (!Profiler::g_ProfileBlocks)
+    return;
 
   // Fetch the current counter register
-  CNTVCT(XB);
+  CNTVCT(X1);
 
-  MOVP2R(XA, &b->profile_data);
+  MOVP2R(X0, &b->profile_data);
 
-  LDR(INDEX_UNSIGNED, XC, XA, offsetof(JitBlock::ProfileData, ticStart));
-  SUB(XB, XB, XC);
+  LDR(INDEX_UNSIGNED, X2, X0, offsetof(JitBlock::ProfileData, ticStart));
+  SUB(X1, X1, X2);
 
   // loads ticCounter and downcountCounter
-  LDP(INDEX_SIGNED, XC, XD, XA, offsetof(JitBlock::ProfileData, ticCounter));
-  ADD(XC, XC, XB);
-  ADDI2R(XD, XD, js.downcountAmount);
+  LDP(INDEX_SIGNED, X2, X3, X0, offsetof(JitBlock::ProfileData, ticCounter));
+  ADD(X2, X2, X1);
+  ADDI2R(X3, X3, js.downcountAmount, X1);
 
   // stores ticCounter and downcountCounter
-  STP(INDEX_SIGNED, XC, XD, XA, offsetof(JitBlock::ProfileData, ticCounter));
-
-  gpr.Unlock(WA, WB, WC, WD);
+  STP(INDEX_SIGNED, X2, X3, X0, offsetof(JitBlock::ProfileData, ticCounter));
 }
 
 void JitArm64::Run()
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
index 82325b0f49..da49e0a3bf 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
@@ -67,8 +67,8 @@ void JitArm64::rfi(UGeckoInstruction inst)
   LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(spr[SPR_SRR0]));
   gpr.Unlock(WB, WC);
 
-  // WA is unlocked in this function
   WriteExceptionExit(WA);
+  gpr.Unlock(WA);
 }
 
 void JitArm64::bx(UGeckoInstruction inst)
@@ -220,6 +220,8 @@ void JitArm64::bcctrx(UGeckoInstruction inst)
   AND(WA, WA, 30, 29);  // Wipe the bottom 2 bits.
 
   WriteExit(WA, inst.LK_3, js.compilerPC + 4);
+
+  gpr.Unlock(WA);
 }
 
 void JitArm64::bclrx(UGeckoInstruction inst)
@@ -275,6 +277,8 @@ void JitArm64::bclrx(UGeckoInstruction inst)
 
   WriteBLRExit(WA);
 
+  gpr.Unlock(WA);
+
   if (conditional)
     SwitchToNearCode();