Merge pull request #9957 from JosJuice/dcbx-faster

Jit: Perform BAT lookup in dcbf/dcbi/dcbst

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

@@ -234,14 +234,41 @@ void Jit64::dcbx(UGeckoInstruction inst)
   JITDISABLE(bJITLoadStoreOff);
 
   X64Reg addr = RSCRATCH;
+  X64Reg value = RSCRATCH2;
   RCOpArg Ra = inst.RA ? gpr.Use(inst.RA, RCMode::Read) : RCOpArg::Imm32(0);
   RCOpArg Rb = gpr.Use(inst.RB, RCMode::Read);
-  RegCache::Realize(Ra, Rb);
+  RCX64Reg tmp = gpr.Scratch();
+  RegCache::Realize(Ra, Rb, tmp);
 
-  MOV_sum(32, addr, Ra, Rb);
+  // Translate effective address to physical address.
-  AND(32, R(addr), Imm8(~31));
+  MOV_sum(32, value, Ra, Rb);
+  FixupBranch bat_lookup_failed;
+  if (MSR.IR)
+  {
+    MOV(32, R(addr), R(value));
+    bat_lookup_failed = BATAddressLookup(value, tmp);
+    AND(32, R(addr), Imm32(0x0001ffff));
+    AND(32, R(value), Imm32(0xfffe0000));
+    OR(32, R(value), R(addr));
+  }
+  MOV(32, R(addr), R(value));
+
+  // Check whether a JIT cache line needs to be invalidated.
+  SHR(32, R(value), Imm8(5 + 5));  // >> 5 for cache line size, >> 5 for width of bitset
+  MOV(64, R(tmp), ImmPtr(GetBlockCache()->GetBlockBitSet()));
+  MOV(32, R(value), MComplex(tmp, value, SCALE_4, 0));
+  SHR(32, R(addr), Imm8(5));
+  BT(32, R(value), R(addr));
+  FixupBranch invalidate_needed = J_CC(CC_C, true);
+
+  SwitchToFarCode();
+  SetJumpTarget(invalidate_needed);
+  SHL(32, R(addr), Imm8(5));
+  if (MSR.IR)
+    SetJumpTarget(bat_lookup_failed);
 
   BitSet32 registersInUse = CallerSavedRegistersInUse();
+  registersInUse[X64Reg(tmp)] = false;
   ABI_PushRegistersAndAdjustStack(registersInUse, 0);
   MOV(32, R(ABI_PARAM1), R(addr));
   MOV(32, R(ABI_PARAM2), Imm32(32));
@@ -249,6 +276,10 @@ void Jit64::dcbx(UGeckoInstruction inst)
   ABI_CallFunction(JitInterface::InvalidateICache);
   ABI_PopRegistersAndAdjustStack(registersInUse, 0);
   asm_routines.ResetStack(*this);
+
+  FixupBranch done = J(true);
+  SwitchToNearCode();
+  SetJumpTarget(done);
 }
 
 void Jit64::dcbt(UGeckoInstruction inst)

Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp

@@ -91,6 +91,16 @@ void EmuCodeBlock::SwitchToNearCode()
   SetCodePtr(m_near_code, m_near_code_end, m_near_code_write_failed);
 }
 
+FixupBranch EmuCodeBlock::BATAddressLookup(X64Reg addr, X64Reg tmp)
+{
+  MOV(64, R(tmp), ImmPtr(&PowerPC::dbat_table[0]));
+  SHR(32, R(addr), Imm8(PowerPC::BAT_INDEX_SHIFT));
+  MOV(32, R(addr), MComplex(tmp, addr, SCALE_4, 0));
+  BT(32, R(addr), Imm8(IntLog2(PowerPC::BAT_MAPPED_BIT)));
+
+  return J_CC(CC_Z, m_far_code.Enabled());
+}
+
 FixupBranch EmuCodeBlock::CheckIfSafeAddress(const OpArg& reg_value, X64Reg reg_addr,
                                              BitSet32 registers_in_use)
 {

Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.h

@@ -49,6 +49,10 @@ public:
     return Gen::M(m_const_pool.GetConstant(&value, sizeof(T), N, index));
   }
 
+  // Writes upper 15 bits of physical address to addr and clobbers the lower 17 bits of addr.
+  // Jumps to the returned FixupBranch if lookup fails.
+  Gen::FixupBranch BATAddressLookup(Gen::X64Reg addr, Gen::X64Reg tmp);
+
   Gen::FixupBranch CheckIfSafeAddress(const Gen::OpArg& reg_value, Gen::X64Reg reg_addr,
                                       BitSet32 registers_in_use);
   void UnsafeLoadRegToReg(Gen::X64Reg reg_addr, Gen::X64Reg reg_value, int accessSize,

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

@@ -230,6 +230,10 @@ protected:
   // Loadstore routines
   void SafeLoadToReg(u32 dest, s32 addr, s32 offsetReg, u32 flags, s32 offset, bool update);
   void SafeStoreFromReg(s32 dest, u32 value, s32 regOffset, u32 flags, s32 offset);
+  // If lookup succeeds, writes upper 15 bits of physical address to addr_out. If not,
+  // jumps to the returned FixupBranch. Clobbers tmp and the 17 lower bits of addr_out.
+  Arm64Gen::FixupBranch BATAddressLookup(Arm64Gen::ARM64Reg addr_out, Arm64Gen::ARM64Reg addr_in,
+                                         Arm64Gen::ARM64Reg tmp);
 
   void DoJit(u32 em_address, JitBlock* b, u32 nextPC);
 

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

@@ -274,6 +274,19 @@ void JitArm64::SafeStoreFromReg(s32 dest, u32 value, s32 regOffset, u32 flags, s
   gpr.Unlock(ARM64Reg::W0, ARM64Reg::W1, ARM64Reg::W30);
 }
 
+FixupBranch JitArm64::BATAddressLookup(ARM64Reg addr_out, ARM64Reg addr_in, ARM64Reg tmp)
+{
+  tmp = EncodeRegTo64(tmp);
+
+  MOVP2R(tmp, PowerPC::dbat_table.data());
+  LSR(addr_out, addr_in, PowerPC::BAT_INDEX_SHIFT);
+  LDR(addr_out, tmp, ArithOption(addr_out, true));
+  FixupBranch pass = TBNZ(addr_out, IntLog2(PowerPC::BAT_MAPPED_BIT));
+  FixupBranch fail = B();
+  SetJumpTarget(pass);
+  return fail;
+}
+
 void JitArm64::lXX(UGeckoInstruction inst)
 {
   INSTRUCTION_START
@@ -539,34 +552,71 @@ void JitArm64::dcbx(UGeckoInstruction inst)
   INSTRUCTION_START
   JITDISABLE(bJITLoadStoreOff);
 
-  gpr.Lock(ARM64Reg::W0);
+  gpr.Lock(ARM64Reg::W0, ARM64Reg::W30);
 
-  ARM64Reg addr = ARM64Reg::W0;
+  ARM64Reg effective_addr = ARM64Reg::W0;
+  ARM64Reg physical_addr = MSR.IR ? gpr.GetReg() : effective_addr;
+  ARM64Reg value = gpr.GetReg();
+  ARM64Reg WA = ARM64Reg::W30;
 
   u32 a = inst.RA, b = inst.RB;
 
   if (a)
-    ADD(addr, gpr.R(a), gpr.R(b));
+    ADD(effective_addr, gpr.R(a), gpr.R(b));
   else
-    MOV(addr, gpr.R(b));
+    MOV(effective_addr, gpr.R(b));
 
-  AND(addr, addr, LogicalImm(~31, 32));  // mask sizeof cacheline
+  // Translate effective address to physical address.
+  FixupBranch bat_lookup_failed;
+  if (MSR.IR)
+  {
+    bat_lookup_failed = BATAddressLookup(physical_addr, effective_addr, WA);
+    BFI(physical_addr, effective_addr, 0, PowerPC::BAT_INDEX_SHIFT);
+  }
+
+  // Check whether a JIT cache line needs to be invalidated.
+  LSR(value, physical_addr, 5 + 5);  // >> 5 for cache line size, >> 5 for width of bitset
+  MOVP2R(EncodeRegTo64(WA), GetBlockCache()->GetBlockBitSet());
+  LDR(value, EncodeRegTo64(WA), ArithOption(EncodeRegTo64(value), true));
+
+  LSR(WA, physical_addr, 5);  // mask sizeof cacheline, & 0x1f is the position within the bitset
+
+  LSRV(value, value, WA);  // move current bit to bit 0
+
+  FixupBranch bit_not_set = TBZ(value, 0);
+  FixupBranch far_addr = B();
+  SwitchToFarCode();
+  SetJumpTarget(far_addr);
+  if (MSR.IR)
+    SetJumpTarget(bat_lookup_failed);
 
   BitSet32 gprs_to_push = gpr.GetCallerSavedUsed();
   BitSet32 fprs_to_push = fpr.GetCallerSavedUsed();
+  gprs_to_push[DecodeReg(effective_addr)] = false;
+  gprs_to_push[DecodeReg(physical_addr)] = false;
+  gprs_to_push[DecodeReg(value)] = false;
+  gprs_to_push[DecodeReg(WA)] = false;
 
   ABI_PushRegisters(gprs_to_push);
   m_float_emit.ABI_PushRegisters(fprs_to_push, ARM64Reg::X30);
 
-  MOVI2R(ARM64Reg::X1, 32);
+  MOVP2R(ARM64Reg::X8, &JitInterface::InvalidateICache);
-  MOVI2R(ARM64Reg::X2, 0);
+  // W0 was already set earlier
-  MOVP2R(ARM64Reg::X3, &JitInterface::InvalidateICache);
+  MOVI2R(ARM64Reg::W1, 32);
-  BLR(ARM64Reg::X3);
+  MOVI2R(ARM64Reg::W2, 0);
+  BLR(ARM64Reg::X8);
 
   m_float_emit.ABI_PopRegisters(fprs_to_push, ARM64Reg::X30);
   ABI_PopRegisters(gprs_to_push);
 
-  gpr.Unlock(ARM64Reg::W0);
+  FixupBranch near_addr = B();
+  SwitchToNearCode();
+  SetJumpTarget(bit_not_set);
+  SetJumpTarget(near_addr);
+
+  gpr.Unlock(effective_addr, value, WA);
+  if (MSR.IR)
+    gpr.Unlock(physical_addr);
 }
 
 void JitArm64::dcbt(UGeckoInstruction inst)

Source/Core/Core/PowerPC/JitCommon/JitCache.cpp

@@ -269,6 +269,11 @@ void JitBaseBlockCache::ErasePhysicalRange(u32 address, u32 length)
   }
 }
 
+u32* JitBaseBlockCache::GetBlockBitSet() const
+{
+  return valid_block.m_valid_block.get();
+}
+
 void JitBaseBlockCache::WriteDestroyBlock(const JitBlock& block)
 {
 }

Source/Core/Core/PowerPC/JitCommon/JitCache.h

@@ -99,18 +99,6 @@ typedef void (*CompiledCode)();
 class ValidBlockBitSet final
 {
 public:
-  ValidBlockBitSet()
-  {
-    m_valid_block.reset(new u32[VALID_BLOCK_ALLOC_ELEMENTS]);
-    ClearAll();
-  }
-
-  void Set(u32 bit) { m_valid_block[bit / 32] |= 1u << (bit % 32); }
-  void Clear(u32 bit) { m_valid_block[bit / 32] &= ~(1u << (bit % 32)); }
-  void ClearAll() { memset(m_valid_block.get(), 0, sizeof(u32) * VALID_BLOCK_ALLOC_ELEMENTS); }
-  bool Test(u32 bit) { return (m_valid_block[bit / 32] & (1u << (bit % 32))) != 0; }
-
-private:
   enum
   {
     // ValidBlockBitSet covers the whole 32-bit address-space in 32-byte
@@ -121,7 +109,19 @@ private:
     // The number of elements in the allocated array. Each u32 contains 32 bits.
     VALID_BLOCK_ALLOC_ELEMENTS = VALID_BLOCK_MASK_SIZE / 32
   };
+  // Directly accessed by Jit64.
   std::unique_ptr<u32[]> m_valid_block;
+
+  ValidBlockBitSet()
+  {
+    m_valid_block.reset(new u32[VALID_BLOCK_ALLOC_ELEMENTS]);
+    ClearAll();
+  }
+
+  void Set(u32 bit) { m_valid_block[bit / 32] |= 1u << (bit % 32); }
+  void Clear(u32 bit) { m_valid_block[bit / 32] &= ~(1u << (bit % 32)); }
+  void ClearAll() { memset(m_valid_block.get(), 0, sizeof(u32) * VALID_BLOCK_ALLOC_ELEMENTS); }
+  bool Test(u32 bit) const { return (m_valid_block[bit / 32] & (1u << (bit % 32))) != 0; }
 };
 
 class JitBaseBlockCache
@@ -163,6 +163,8 @@ public:
   void InvalidateICache(u32 address, u32 length, bool forced);
   void ErasePhysicalRange(u32 address, u32 length);
 
+  u32* GetBlockBitSet() const;
+
 protected:
   virtual void DestroyBlock(JitBlock& block);