Merge pull request #4650 from degasus/jitcache

JitCache: Use pointers in the ICache and dispatchers.

Source/Core/Core/PowerPC/Jit64/JitAsm.cpp

@@ -106,31 +106,24 @@ void Jit64AsmRoutineManager::Generate()
   // The following is a translation of JitBaseBlockCache::Dispatch into assembly.
 
   // Fast block number lookup.
+  // ((PC >> 2) & mask) * sizeof(JitBlock*) = (PC & (mask << 2)) * 2
   MOV(32, R(RSCRATCH), PPCSTATE(pc));
   u64 icache = reinterpret_cast<u64>(g_jit->GetBlockCache()->GetICache());
   AND(32, R(RSCRATCH), Imm32(JitBaseBlockCache::iCache_Mask << 2));
   if (icache <= INT_MAX)
   {
-    MOV(32, R(RSCRATCH), MDisp(RSCRATCH, static_cast<s32>(icache)));
+    MOV(64, R(RSCRATCH), MScaled(RSCRATCH, SCALE_2, static_cast<s32>(icache)));
   }
   else
   {
     MOV(64, R(RSCRATCH2), Imm64(icache));
-    MOV(32, R(RSCRATCH), MRegSum(RSCRATCH2, RSCRATCH));
+    MOV(64, R(RSCRATCH), MComplex(RSCRATCH2, RSCRATCH, SCALE_2, 0));
   }
 
-  // Check whether the block we found matches the current state.
+  // Check if we found a block.
-  u64 blocks = reinterpret_cast<u64>(g_jit->GetBlockCache()->GetBlocks());
+  TEST(64, R(RSCRATCH), R(RSCRATCH));
-  IMUL(32, RSCRATCH, R(RSCRATCH), Imm32(sizeof(JitBlock)));
+  FixupBranch not_found = J_CC(CC_Z);
-  if (blocks <= INT_MAX)
+
-  {
-    ADD(64, R(RSCRATCH), Imm32(static_cast<s32>(blocks)));
-  }
-  else
-  {
-    MOV(64, R(RSCRATCH2), Imm64(blocks));
-    ADD(64, R(RSCRATCH), R(RSCRATCH2));
-  }
   // Check both block.effectiveAddress and block.msrBits.
   MOV(32, R(RSCRATCH2), PPCSTATE(msr));
   AND(32, R(RSCRATCH2), Imm32(JitBlock::JIT_CACHE_MSR_MASK));
@@ -138,10 +131,11 @@ void Jit64AsmRoutineManager::Generate()
   MOV(32, R(RSCRATCH_EXTRA), PPCSTATE(pc));
   OR(64, R(RSCRATCH2), R(RSCRATCH_EXTRA));
   CMP(64, R(RSCRATCH2), MDisp(RSCRATCH, static_cast<s32>(offsetof(JitBlock, effectiveAddress))));
-  FixupBranch notfound = J_CC(CC_NE);
+  FixupBranch state_mismatch = J_CC(CC_NE);
   // Success; branch to the block we found.
   JMPptr(MDisp(RSCRATCH, static_cast<s32>(offsetof(JitBlock, normalEntry))));
-  SetJumpTarget(notfound);
+  SetJumpTarget(not_found);
+  SetJumpTarget(state_mismatch);
 
   // Failure; call into the block cache to update the state, then try again.
   // (We need to loop because Jit() might not actually generate a block

Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp

@@ -73,18 +73,12 @@ void JitArm64::GenerateAsm()
     // iCache[(address >> 2) & iCache_Mask];
     ARM64Reg pc_masked = W25;
     ARM64Reg cache_base = X27;
-    ARM64Reg block_num = W27;
-    ANDI2R(pc_masked, DISPATCHER_PC, JitBaseBlockCache::iCache_Mask << 2);
-    MOVP2R(cache_base, g_jit->GetBlockCache()->GetICache());
-    LDR(block_num, cache_base, EncodeRegTo64(pc_masked));
-
-    // blocks[block_num]
     ARM64Reg block = X30;
-    ARM64Reg jit_block_size = W24;
+    ORRI2R(pc_masked, WZR, JitBaseBlockCache::iCache_Mask << 3);
-    MOVI2R(jit_block_size, sizeof(JitBlock));
+    AND(pc_masked, pc_masked, DISPATCHER_PC, ArithOption(DISPATCHER_PC, ST_LSL, 1));
-    MUL(block_num, block_num, jit_block_size);
+    MOVP2R(cache_base, g_jit->GetBlockCache()->GetICache());
-    MOVP2R(block, g_jit->GetBlockCache()->GetBlocks());
+    LDR(block, cache_base, EncodeRegTo64(pc_masked));
-    ADD(block, block, EncodeRegTo64(block_num));
+    FixupBranch not_found = CBZ(block);
 
     // b.effectiveAddress != addr || b.msrBits != msr
     ARM64Reg pc_and_msr = W25;
@@ -102,6 +96,7 @@ void JitArm64::GenerateAsm()
     // return blocks[block_num].normalEntry;
     LDR(INDEX_UNSIGNED, block, block, offsetof(JitBlock, normalEntry));
     BR(block);
+    SetJumpTarget(not_found);
     SetJumpTarget(pc_missmatch);
     SetJumpTarget(msr_missmatch);
   }

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

@@ -46,7 +46,7 @@ void JitBaseBlockCache::Init()
   s_clear_jit_cache_thread_safe = CoreTiming::RegisterEvent("clearJitCache", ClearCacheThreadSafe);
   JitRegister::Init(SConfig::GetInstance().m_perfDir);
 
-  iCache.fill(0);
+  iCache.fill(nullptr);
   Clear();
 }
 
@@ -99,12 +99,7 @@ bool JitBaseBlockCache::IsFull() const
   return num_blocks >= MAX_NUM_BLOCKS - 1;
 }
 
-JitBlock* JitBaseBlockCache::GetBlocks()
+JitBlock** JitBaseBlockCache::GetICache()
-{
-  return blocks.data();
-}
-
-int* JitBaseBlockCache::GetICache()
 {
   return iCache.data();
 }
@@ -140,9 +135,8 @@ void JitBaseBlockCache::FinalizeBlock(JitBlock& block, bool block_link, const u8
         std::make_pair(old_b.physicalAddress + 4 * old_b.originalSize - 1, old_b.physicalAddress));
     DestroyBlock(old_b, true);
   }
-  const int block_num = static_cast<int>(&block - &blocks[0]);
   start_block_map[block.physicalAddress] = &block;
-  FastLookupEntryForAddress(block.effectiveAddress) = block_num;
+  FastLookupEntryForAddress(block.effectiveAddress) = &block;
 
   u32 pAddr = block.physicalAddress;
 
@@ -190,16 +184,16 @@ JitBlock* JitBaseBlockCache::GetBlockFromStartAddress(u32 addr, u32 msr)
 
 const u8* JitBaseBlockCache::Dispatch()
 {
-  int block_num = FastLookupEntryForAddress(PC);
+  JitBlock* block = FastLookupEntryForAddress(PC);
 
-  while (blocks[block_num].effectiveAddress != PC ||
+  while (!block || block->effectiveAddress != PC ||
-         blocks[block_num].msrBits != (MSR & JitBlock::JIT_CACHE_MSR_MASK))
+         block->msrBits != (MSR & JitBlock::JIT_CACHE_MSR_MASK))
   {
     MoveBlockIntoFastCache(PC, MSR & JitBlock::JIT_CACHE_MSR_MASK);
-    block_num = FastLookupEntryForAddress(PC);
+    block = FastLookupEntryForAddress(PC);
   }
 
-  return blocks[block_num].normalEntry;
+  return block->normalEntry;
 }
 
 void JitBaseBlockCache::InvalidateICache(u32 address, const u32 length, bool forced)
@@ -326,7 +320,7 @@ void JitBaseBlockCache::DestroyBlock(JitBlock& block, bool invalidate)
   }
   block.invalid = true;
   start_block_map.erase(block.physicalAddress);
-  FastLookupEntryForAddress(block.effectiveAddress) = 0;
+  FastLookupEntryForAddress(block.effectiveAddress) = nullptr;
 
   UnlinkBlock(block);
 
@@ -353,12 +347,12 @@ void JitBaseBlockCache::MoveBlockIntoFastCache(u32 addr, u32 msr)
   }
   else
   {
-    FastLookupEntryForAddress(addr) = static_cast<int>(block - &blocks[0]);
+    FastLookupEntryForAddress(addr) = block;
     LinkBlock(*block);
   }
 }
 
-int& JitBaseBlockCache::FastLookupEntryForAddress(u32 address)
+JitBlock*& JitBaseBlockCache::FastLookupEntryForAddress(u32 address)
 {
   return iCache[(address >> 2) & iCache_Mask];
 }

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

@@ -127,8 +127,7 @@ public:
   bool IsFull() const;
 
   // Code Cache
-  JitBlock* GetBlocks();
+  JitBlock** GetICache();
-  int* GetICache();
   void RunOnBlocks(std::function<void(const JitBlock&)> f);
 
   JitBlock* AllocateBlock(u32 em_address);
@@ -164,7 +163,7 @@ private:
   void MoveBlockIntoFastCache(u32 em_address, u32 msr);
 
   // Fast but risky block lookup based on iCache.
-  int& FastLookupEntryForAddress(u32 address);
+  JitBlock*& FastLookupEntryForAddress(u32 address);
 
   // We store the metadata of all blocks in a linear way within this array.
   // Note: blocks[0] must not be used as it is referenced as invalid block in iCache.
@@ -190,5 +189,5 @@ private:
 
   // This array is indexed with the masked PC and likely holds the correct block id.
   // This is used as a fast cache of start_block_map used in the assembly dispatcher.
-  std::array<int, iCache_Num_Elements> iCache;  // start_addr & mask -> number
+  std::array<JitBlock*, iCache_Num_Elements> iCache;  // start_addr & mask -> number
 };