Merge pull request #9770 from JosJuice/jits-accidental-gt

Jits: Fix accidentally setting GT in CR when clearing EQ

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

@@ -116,6 +116,7 @@ public:
   void SetCRFieldBit(int field, int bit, Gen::X64Reg in);
   void ClearCRFieldBit(int field, int bit);
   void SetCRFieldBit(int field, int bit);
+  void FixGTBeforeSettingCRFieldBit(Gen::X64Reg reg);
 
   // Generates a branch that will check if a given bit of a CR register part
   // is set or not.

Source/Core/Core/PowerPC/Jit64/Jit_SystemRegisters.cpp

@@ -54,16 +54,8 @@ void Jit64::SetCRFieldBit(int field, int bit, X64Reg in)
   MOV(64, R(RSCRATCH2), CROffset(field));
   MOVZX(32, 8, in, R(in));
 
-  // Gross but necessary; if the input is totally zero and we set SO or LT,
-  // or even just add the (1<<32), GT will suddenly end up set without us
-  // intending to. This can break actual games, so fix it up.
   if (bit != PowerPC::CR_GT_BIT)
-  {
+    FixGTBeforeSettingCRFieldBit(RSCRATCH2);
-    TEST(64, R(RSCRATCH2), R(RSCRATCH2));
-    FixupBranch dont_clear_gt = J_CC(CC_NZ);
-    BTS(64, R(RSCRATCH2), Imm8(63));
-    SetJumpTarget(dont_clear_gt);
-  }
 
   switch (bit)
   {
@@ -107,7 +99,10 @@ void Jit64::ClearCRFieldBit(int field, int bit)
     break;
 
   case PowerPC::CR_EQ_BIT:
-    OR(64, CROffset(field), Imm8(1));
+    MOV(64, R(RSCRATCH), CROffset(field));
+    FixGTBeforeSettingCRFieldBit(RSCRATCH);
+    OR(64, R(RSCRATCH), Imm8(1));
+    MOV(64, CROffset(field), R(RSCRATCH));
     break;
 
   case PowerPC::CR_GT_BIT:
@@ -126,12 +121,7 @@ void Jit64::SetCRFieldBit(int field, int bit)
 {
   MOV(64, R(RSCRATCH), CROffset(field));
   if (bit != PowerPC::CR_GT_BIT)
-  {
+    FixGTBeforeSettingCRFieldBit(RSCRATCH);
-    TEST(64, R(RSCRATCH), R(RSCRATCH));
-    FixupBranch dont_clear_gt = J_CC(CC_NZ);
-    BTS(64, R(RSCRATCH), Imm8(63));
-    SetJumpTarget(dont_clear_gt);
-  }
 
   switch (bit)
   {
@@ -157,6 +147,17 @@ void Jit64::SetCRFieldBit(int field, int bit)
   MOV(64, CROffset(field), R(RSCRATCH));
 }
 
+void Jit64::FixGTBeforeSettingCRFieldBit(Gen::X64Reg reg)
+{
+  // Gross but necessary; if the input is totally zero and we set SO or LT,
+  // or even just add the (1<<32), GT will suddenly end up set without us
+  // intending to. This can break actual games, so fix it up.
+  TEST(64, R(reg), R(reg));
+  FixupBranch dont_clear_gt = J_CC(CC_NZ);
+  BTS(64, R(reg), Imm8(63));
+  SetJumpTarget(dont_clear_gt);
+}
+
 FixupBranch Jit64::JumpIfCRFieldBit(int field, int bit, bool jump_if_set)
 {
   switch (bit)

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

@@ -255,6 +255,7 @@ protected:
   void FakeLKExit(u32 exit_address_after_return);
   void WriteBLRExit(Arm64Gen::ARM64Reg dest);
 
+  void FixGTBeforeSettingCRFieldBit(Arm64Gen::ARM64Reg reg);
   Arm64Gen::FixupBranch JumpIfCRFieldBit(int field, int bit, bool jump_if_set);
 
   void ComputeRC0(Arm64Gen::ARM64Reg reg);

Source/Core/Core/PowerPC/JitArm64/JitArm64_SystemRegisters.cpp

@@ -36,6 +36,19 @@ FixupBranch JitArm64::JumpIfCRFieldBit(int field, int bit, bool jump_if_set)
   }
 }
 
+void JitArm64::FixGTBeforeSettingCRFieldBit(Arm64Gen::ARM64Reg reg)
+{
+  // Gross but necessary; if the input is totally zero and we set SO or LT,
+  // or even just add the (1<<32), GT will suddenly end up set without us
+  // intending to. This can break actual games, so fix it up.
+  ARM64Reg WA = gpr.GetReg();
+  ARM64Reg XA = EncodeRegTo64(WA);
+  ORR(XA, reg, 64 - 63, 0, true);  // XB | 1<<63
+  CMP(reg, ARM64Reg::ZR);
+  CSEL(reg, reg, XA, CC_NEQ);
+  gpr.Unlock(WA);
+}
+
 void JitArm64::mtmsr(UGeckoInstruction inst)
 {
   INSTRUCTION_START
@@ -432,6 +445,7 @@ void JitArm64::crXXX(UGeckoInstruction inst)
       break;
 
     case PowerPC::CR_EQ_BIT:
+      FixGTBeforeSettingCRFieldBit(XA);
       ORR(XA, XA, 0, 0, true);  // XA | 1<<0
       break;
 
@@ -457,14 +471,7 @@ void JitArm64::crXXX(UGeckoInstruction inst)
     ARM64Reg XA = gpr.CR(field);
 
     if (bit != PowerPC::CR_GT_BIT)
-    {
+      FixGTBeforeSettingCRFieldBit(XA);
-      ARM64Reg WB = gpr.GetReg();
-      ARM64Reg XB = EncodeRegTo64(WB);
-      ORR(XB, XA, 64 - 63, 0, true);  // XA | 1<<63
-      CMP(XA, ARM64Reg::ZR);
-      CSEL(XA, XA, XB, CC_NEQ);
-      gpr.Unlock(WB);
-    }
 
     switch (bit)
     {
@@ -569,18 +576,8 @@ void JitArm64::crXXX(UGeckoInstruction inst)
   gpr.BindCRToRegister(field, true);
   XB = gpr.CR(field);
 
-  // Gross but necessary; if the input is totally zero and we set SO or LT,
-  // or even just add the (1<<32), GT will suddenly end up set without us
-  // intending to. This can break actual games, so fix it up.
   if (bit != PowerPC::CR_GT_BIT)
-  {
+    FixGTBeforeSettingCRFieldBit(XB);
-    ARM64Reg WC = gpr.GetReg();
-    ARM64Reg XC = EncodeRegTo64(WC);
-    ORR(XC, XB, 64 - 63, 0, true);  // XB | 1<<63
-    CMP(XB, ARM64Reg::ZR);
-    CSEL(XB, XB, XC, CC_NEQ);
-    gpr.Unlock(WC);
-  }
 
   switch (bit)
   {