Merge pull request #9593 from JosJuice/jitarm64-constant-carry

JitArm64: Constant carry flag optimizations

Source/Core/Common/Arm64Emitter.h

@@ -673,6 +673,7 @@ public:
     CSINV(Rd, zr, zr, (CCFlags)((u32)cond ^ 1));
   }
   void NEG(ARM64Reg Rd, ARM64Reg Rs) { SUB(Rd, Is64Bit(Rd) ? ARM64Reg::ZR : ARM64Reg::WZR, Rs); }
+  void NEGS(ARM64Reg Rd, ARM64Reg Rs) { SUBS(Rd, Is64Bit(Rd) ? ARM64Reg::ZR : ARM64Reg::WZR, Rs); }
   // Data-Processing 1 source
   void RBIT(ARM64Reg Rd, ARM64Reg Rn);
   void REV16(ARM64Reg Rd, ARM64Reg Rn);

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

@@ -926,8 +926,7 @@ bool Jit64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
 
   js.downcountAmount = 0;
   js.skipInstructions = 0;
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
-  js.carryFlagInverted = false;
   js.constantGqr.clear();
 
   // Assume that GQR values don't change often at runtime. Many paired-heavy games use largely float

Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp

@@ -63,26 +63,29 @@ void Jit64::GenerateOverflow()
 
 void Jit64::FinalizeCarry(CCFlags cond)
 {
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
-  js.carryFlagInverted = false;
   if (js.op->wantsCA)
   {
     // Not actually merging instructions, but the effect is equivalent (we can't have
     // breakpoints/etc in between).
     if (CanMergeNextInstructions(1) && js.op[1].wantsCAInFlags)
     {
-      if (cond == CC_C || cond == CC_NC)
+      if (cond == CC_C)
       {
-        js.carryFlagInverted = cond == CC_NC;
+        js.carryFlag = CarryFlag::InHostCarry;
+      }
+      else if (cond == CC_NC)
+      {
+        js.carryFlag = CarryFlag::InHostCarryInverted;
       }
       else
       {
         // convert the condition to a carry flag (is there a better way?)
         SETcc(cond, R(RSCRATCH));
         SHR(8, R(RSCRATCH), Imm8(1));
+        js.carryFlag = CarryFlag::InHostCarry;
       }
       LockFlags();
-      js.carryFlagSet = true;
     }
     else
     {
@@ -94,8 +97,7 @@ void Jit64::FinalizeCarry(CCFlags cond)
 // Unconditional version
 void Jit64::FinalizeCarry(bool ca)
 {
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
-  js.carryFlagInverted = false;
   if (js.op->wantsCA)
   {
     if (CanMergeNextInstructions(1) && js.op[1].wantsCAInFlags)
@@ -105,7 +107,7 @@ void Jit64::FinalizeCarry(bool ca)
       else
         CLC();
       LockFlags();
-      js.carryFlagSet = true;
+      js.carryFlag = CarryFlag::InHostCarry;
     }
     else if (ca)
     {
@@ -1475,7 +1477,7 @@ void Jit64::arithXex(UGeckoInstruction inst)
   int d = inst.RD;
   bool same_input_sub = !add && regsource && a == b;
 
-  if (!js.carryFlagSet)
+  if (js.carryFlag == CarryFlag::InPPCState)
     JitGetAndClearCAOV(inst.OE);
   else
     UnlockFlags();
@@ -1488,7 +1490,7 @@ void Jit64::arithXex(UGeckoInstruction inst)
     RegCache::Realize(Rd);
 
     // Convert carry to borrow
-    if (!js.carryFlagInverted)
+    if (js.carryFlag != CarryFlag::InHostCarryInverted)
       CMC();
     SBB(32, Rd, Rd);
     invertedCarry = true;
@@ -1499,7 +1501,7 @@ void Jit64::arithXex(UGeckoInstruction inst)
     RCX64Reg Rd = gpr.Bind(d, RCMode::ReadWrite);
     RegCache::Realize(Ra, Rd);
 
-    if (!js.carryFlagInverted)
+    if (js.carryFlag != CarryFlag::InHostCarryInverted)
       CMC();
     SBB(32, Rd, Ra);
     invertedCarry = true;
@@ -1519,14 +1521,14 @@ void Jit64::arithXex(UGeckoInstruction inst)
       NOT(32, Rd);
     // if the source is an immediate, we can invert carry by going from add -> sub and doing src =
     // -1 - src
-    if (js.carryFlagInverted && source.IsImm())
+    if (js.carryFlag == CarryFlag::InHostCarryInverted && source.IsImm())
     {
       SBB(32, Rd, Imm32(-1 - source.SImm32()));
       invertedCarry = true;
     }
     else
     {
-      if (js.carryFlagInverted)
+      if (js.carryFlag == CarryFlag::InHostCarryInverted)
         CMC();
       ADC(32, Rd, source);
     }

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

@@ -638,7 +638,7 @@ void JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
   js.downcountAmount = 0;
   js.skipInstructions = 0;
   js.curBlock = b;
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
   js.numLoadStoreInst = 0;
   js.numFloatingPointInst = 0;
 

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

@@ -233,7 +233,7 @@ private:
   void ComputeRC0(Arm64Gen::ARM64Reg reg);
   void ComputeRC0(u64 imm);
   void ComputeCarry(Arm64Gen::ARM64Reg reg);  // reg must contain 0 or 1
-  void ComputeCarry(bool Carry);
+  void ComputeCarry(bool carry);
   void ComputeCarry();
   void FlushCarry();
 

Source/Core/Core/PowerPC/JitArm64/JitArm64_Integer.cpp

@@ -32,7 +32,7 @@ void JitArm64::ComputeRC0(u64 imm)
 
 void JitArm64::ComputeCarry(ARM64Reg reg)
 {
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
 
   if (!js.op->wantsCA)
     return;
@@ -40,7 +40,7 @@ void JitArm64::ComputeCarry(ARM64Reg reg)
   if (CanMergeNextInstructions(1) && js.op[1].wantsCAInFlags)
   {
     CMP(reg, 1);
-    js.carryFlagSet = true;
+    js.carryFlag = CarryFlag::InHostCarry;
   }
   else
   {
@@ -48,52 +48,57 @@ void JitArm64::ComputeCarry(ARM64Reg reg)
   }
 }
 
-void JitArm64::ComputeCarry(bool Carry)
+void JitArm64::ComputeCarry(bool carry)
 {
-  js.carryFlagSet = false;
+  js.carryFlag = carry ? CarryFlag::ConstantTrue : CarryFlag::ConstantFalse;
-
-  if (!js.op->wantsCA)
-    return;
-
-  if (Carry)
-  {
-    ARM64Reg WA = gpr.GetReg();
-    MOVI2R(WA, 1);
-    STRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
-    gpr.Unlock(WA);
-    return;
-  }
-
-  STRB(IndexType::Unsigned, ARM64Reg::WSP, PPC_REG, PPCSTATE_OFF(xer_ca));
 }
 
 void JitArm64::ComputeCarry()
 {
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
 
   if (!js.op->wantsCA)
     return;
 
-  js.carryFlagSet = true;
+  js.carryFlag = CarryFlag::InHostCarry;
   if (CanMergeNextInstructions(1) && js.op[1].opinfo->type == ::OpType::Integer)
-  {
     return;
-  }
 
   FlushCarry();
 }
 
 void JitArm64::FlushCarry()
 {
-  if (!js.carryFlagSet)
+  switch (js.carryFlag)
-    return;
+  {
-
+  case CarryFlag::InPPCState:
+  {
+    break;
+  }
+  case CarryFlag::InHostCarry:
+  {
     ARM64Reg WA = gpr.GetReg();
-  CSINC(WA, ARM64Reg::WSP, ARM64Reg::WSP, CC_CC);
+    CSET(WA, CC_CS);
     STRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
     gpr.Unlock(WA);
+    break;
+  }
+  case CarryFlag::ConstantTrue:
+  {
+    ARM64Reg WA = gpr.GetReg();
+    MOVI2R(WA, 1);
+    STRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
+    gpr.Unlock(WA);
+    break;
+  }
+  case CarryFlag::ConstantFalse:
+  {
+    STRB(IndexType::Unsigned, ARM64Reg::WZR, PPC_REG, PPCSTATE_OFF(xer_ca));
+    break;
+  }
+  }
 
-  js.carryFlagSet = false;
+  js.carryFlag = CarryFlag::InPPCState;
 }
 
 void JitArm64::reg_imm(u32 d, u32 a, u32 value, u32 (*do_op)(u32, u32),
@@ -690,7 +695,7 @@ void JitArm64::srawix(UGeckoInstruction inst)
       else
       {
         CSINC(WA, ARM64Reg::WSP, ARM64Reg::WSP, CC_EQ);
-        STRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
+        ComputeCarry(WA);
       }
       gpr.Unlock(WA);
     }
@@ -841,27 +846,49 @@ void JitArm64::addzex(UGeckoInstruction inst)
 
   int a = inst.RA, d = inst.RD;
 
-  if (js.carryFlagSet)
+  switch (js.carryFlag)
+  {
+  case CarryFlag::InPPCState:
+  {
+    gpr.BindToRegister(d, d == a);
+    ARM64Reg WA = d == a ? gpr.GetReg() : gpr.R(d);
+
+    LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
+    ADDS(gpr.R(d), gpr.R(a), WA);
+    ComputeCarry();
+
+    if (d == a)
+      gpr.Unlock(WA);
+
+    break;
+  }
+  case CarryFlag::InHostCarry:
   {
     gpr.BindToRegister(d, d == a);
     ADCS(gpr.R(d), gpr.R(a), ARM64Reg::WZR);
+    ComputeCarry();
+    break;
   }
-  else if (d == a)
+  case CarryFlag::ConstantTrue:
   {
-    gpr.BindToRegister(d, true);
+    gpr.BindToRegister(d, d == a);
-    ARM64Reg WA = gpr.GetReg();
+    ADDS(gpr.R(d), gpr.R(a), 1);
-    LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
+    ComputeCarry();
-    ADDS(gpr.R(d), gpr.R(a), WA);
+    break;
-    gpr.Unlock(WA);
   }
-  else
+  case CarryFlag::ConstantFalse:
+  {
+    if (d != a)
     {
       gpr.BindToRegister(d, false);
-    LDRB(IndexType::Unsigned, gpr.R(d), PPC_REG, PPCSTATE_OFF(xer_ca));
+      MOV(gpr.R(d), gpr.R(a));
-    ADDS(gpr.R(d), gpr.R(a), gpr.R(d));
+    }
+
+    ComputeCarry(false);
+    break;
+  }
   }
 
-  ComputeCarry();
   if (inst.Rc)
     ComputeRC0(gpr.R(d));
 }
@@ -909,21 +936,39 @@ void JitArm64::subfex(UGeckoInstruction inst)
     u32 i = gpr.GetImm(a), j = gpr.GetImm(b);
 
     gpr.BindToRegister(d, false);
+
+    switch (js.carryFlag)
+    {
+    case CarryFlag::InPPCState:
+    {
       ARM64Reg WA = gpr.GetReg();
-    if (js.carryFlagSet)
-    {
-      MOVI2R(WA, ~i + j);
-      ADC(gpr.R(d), WA, ARM64Reg::WZR);
-    }
-    else
-    {
       LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
       ADDI2R(gpr.R(d), WA, ~i + j, gpr.R(d));
-    }
       gpr.Unlock(WA);
+      break;
+    }
+    case CarryFlag::InHostCarry:
+    {
+      ARM64Reg WA = gpr.GetReg();
+      MOVI2R(WA, ~i + j);
+      ADC(gpr.R(d), WA, ARM64Reg::WZR);
+      gpr.Unlock(WA);
+      break;
+    }
+    case CarryFlag::ConstantTrue:
+    {
+      gpr.SetImmediate(d, ~i + j + 1);
+      break;
+    }
+    case CarryFlag::ConstantFalse:
+    {
+      gpr.SetImmediate(d, ~i + j);
+      break;
+    }
+    }
 
-    bool must_have_carry = Interpreter::Helper_Carry(~i, j);
+    const bool must_have_carry = Interpreter::Helper_Carry(~i, j);
-    bool might_have_carry = (~i + j) == 0xFFFFFFFF;
+    const bool might_have_carry = (~i + j) == 0xFFFFFFFF;
 
     if (must_have_carry)
     {
@@ -931,7 +976,7 @@ void JitArm64::subfex(UGeckoInstruction inst)
     }
     else if (might_have_carry)
     {
-      // carry stay as it is
+      // carry stays as it is
     }
     else
     {
@@ -940,26 +985,50 @@ void JitArm64::subfex(UGeckoInstruction inst)
   }
   else
   {
-    ARM64Reg WA = gpr.GetReg();
+    ARM64Reg WA = js.carryFlag != CarryFlag::ConstantTrue ? gpr.GetReg() : ARM64Reg::INVALID_REG;
     gpr.BindToRegister(d, d == a || d == b);
 
-    // upload the carry state
+    switch (js.carryFlag)
-    if (!js.carryFlagSet)
     {
+    case CarryFlag::InPPCState:
+    {
+      // upload the carry state
       LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
       CMP(WA, 1);
+      [[fallthrough]];
     }
-
+    case CarryFlag::InHostCarry:
-    // d = ~a + b + carry;
+    {
       if (gpr.IsImm(a))
         MOVI2R(WA, u32(~gpr.GetImm(a)));
       else
         MVN(WA, gpr.R(a));
+
       ADCS(gpr.R(d), WA, gpr.R(b));
-
-    gpr.Unlock(WA);
-
       ComputeCarry();
+      break;
+    }
+    case CarryFlag::ConstantTrue:
+    {
+      SUBS(gpr.R(d), gpr.R(b), gpr.R(a));
+      ComputeCarry();
+      break;
+    }
+    case CarryFlag::ConstantFalse:
+    {
+      if (gpr.IsImm(a))
+        MOVI2R(WA, u32(~gpr.GetImm(a)));
+      else
+        MVN(WA, gpr.R(a));
+
+      ADDS(gpr.R(d), WA, gpr.R(b));
+      ComputeCarry();
+      break;
+    }
+    }
+
+    if (WA != ARM64Reg::INVALID_REG)
+      gpr.Unlock(WA);
   }
 
   if (inst.Rc)
@@ -1008,21 +1077,38 @@ void JitArm64::subfzex(UGeckoInstruction inst)
 
   gpr.BindToRegister(d, d == a);
 
-  if (js.carryFlagSet)
+  switch (js.carryFlag)
   {
-    MVN(gpr.R(d), gpr.R(a));
+  case CarryFlag::InPPCState:
-    ADCS(gpr.R(d), gpr.R(d), ARM64Reg::WZR);
-  }
-  else
   {
     ARM64Reg WA = gpr.GetReg();
     LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
     MVN(gpr.R(d), gpr.R(a));
     ADDS(gpr.R(d), gpr.R(d), WA);
-    gpr.Unlock(WA);
-  }
-
     ComputeCarry();
+    gpr.Unlock(WA);
+    break;
+  }
+  case CarryFlag::InHostCarry:
+  {
+    MVN(gpr.R(d), gpr.R(a));
+    ADCS(gpr.R(d), gpr.R(d), ARM64Reg::WZR);
+    ComputeCarry();
+    break;
+  }
+  case CarryFlag::ConstantTrue:
+  {
+    NEGS(gpr.R(d), gpr.R(a));
+    ComputeCarry();
+    break;
+  }
+  case CarryFlag::ConstantFalse:
+  {
+    MVN(gpr.R(d), gpr.R(a));
+    ComputeCarry(false);
+    break;
+  }
+  }
 
   if (inst.Rc)
     ComputeRC0(gpr.R(d));
@@ -1070,21 +1156,39 @@ void JitArm64::addex(UGeckoInstruction inst)
     u32 i = gpr.GetImm(a), j = gpr.GetImm(b);
 
     gpr.BindToRegister(d, false);
+
+    switch (js.carryFlag)
+    {
+    case CarryFlag::InPPCState:
+    {
       ARM64Reg WA = gpr.GetReg();
-    if (js.carryFlagSet)
-    {
-      MOVI2R(WA, i + j);
-      ADC(gpr.R(d), WA, ARM64Reg::WZR);
-    }
-    else
-    {
       LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
       ADDI2R(gpr.R(d), WA, i + j, gpr.R(d));
-    }
       gpr.Unlock(WA);
+      break;
+    }
+    case CarryFlag::InHostCarry:
+    {
+      ARM64Reg WA = gpr.GetReg();
+      MOVI2R(WA, i + j);
+      ADC(gpr.R(d), WA, ARM64Reg::WZR);
+      gpr.Unlock(WA);
+      break;
+    }
+    case CarryFlag::ConstantTrue:
+    {
+      gpr.SetImmediate(d, i + j + 1);
+      break;
+    }
+    case CarryFlag::ConstantFalse:
+    {
+      gpr.SetImmediate(d, i + j);
+      break;
+    }
+    }
 
-    bool must_have_carry = Interpreter::Helper_Carry(i, j);
+    const bool must_have_carry = Interpreter::Helper_Carry(i, j);
-    bool might_have_carry = (i + j) == 0xFFFFFFFF;
+    const bool might_have_carry = (i + j) == 0xFFFFFFFF;
 
     if (must_have_carry)
     {
@@ -1092,7 +1196,7 @@ void JitArm64::addex(UGeckoInstruction inst)
     }
     else if (might_have_carry)
     {
-      // carry stay as it is
+      // carry stays as it is
     }
     else
     {
@@ -1103,19 +1207,60 @@ void JitArm64::addex(UGeckoInstruction inst)
   {
     gpr.BindToRegister(d, d == a || d == b);
 
-    // upload the carry state
+    if (js.carryFlag == CarryFlag::ConstantTrue && !gpr.IsImm(a) && !gpr.IsImm(b))
-    if (!js.carryFlagSet)
     {
+      CMP(ARM64Reg::WZR, ARM64Reg::WZR);
+      js.carryFlag = CarryFlag::InHostCarry;
+    }
+
+    switch (js.carryFlag)
+    {
+    case CarryFlag::InPPCState:
+    {
+      // upload the carry state
       ARM64Reg WA = gpr.GetReg();
       LDRB(IndexType::Unsigned, WA, PPC_REG, PPCSTATE_OFF(xer_ca));
       CMP(WA, 1);
       gpr.Unlock(WA);
+      [[fallthrough]];
     }
-
+    case CarryFlag::InHostCarry:
-    // d = a + b + carry;
+    {
       ADCS(gpr.R(d), gpr.R(a), gpr.R(b));
-
       ComputeCarry();
+      break;
+    }
+    case CarryFlag::ConstantTrue:
+    {
+      if (!gpr.IsImm(b))
+        std::swap(a, b);
+      ASSERT(gpr.IsImm(b));
+
+      ARM64Reg WA = gpr.GetReg();
+      const u32 imm = gpr.GetImm(b) + 1;
+      if (imm == 0)
+      {
+        if (d != a)
+          MOV(gpr.R(d), gpr.R(a));
+
+        ComputeCarry(true);
+      }
+      else
+      {
+        ADDSI2R(gpr.R(d), gpr.R(a), imm, WA);
+        ComputeCarry();
+      }
+      gpr.Unlock(WA);
+
+      break;
+    }
+    case CarryFlag::ConstantFalse:
+    {
+      ADDS(gpr.R(d), gpr.R(a), gpr.R(b));
+      ComputeCarry();
+      break;
+    }
+    }
   }
 
   if (inst.Rc)

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

@@ -42,6 +42,19 @@
 class JitBase : public CPUCoreBase
 {
 protected:
+  enum class CarryFlag
+  {
+    InPPCState,
+    InHostCarry,
+#ifdef _M_X86_64
+    InHostCarryInverted,
+#endif
+#ifdef _M_ARM_64
+    ConstantTrue,
+    ConstantFalse,
+#endif
+  };
+
   struct JitOptions
   {
     bool enableBlocklink;
@@ -73,8 +86,7 @@ protected:
     bool firstFPInstructionFound;
     bool isLastInstruction;
     int skipInstructions;
-    bool carryFlagSet;
+    CarryFlag carryFlag;
-    bool carryFlagInverted;
 
     bool generatingTrampoline = false;
     u8* trampolineExceptionHandler;