Make memory breakpoint faster

Currently, slowmem is used at any time that memory breakpoints are in use.  This commit makes it so that whenever the DBAT gets updated, if the address is overllaping any memchecks, it forces the use of slowmem.  This allows to keep fastmem for any other cases and noticably increases performance when using memory breakpoints.

Source/Core/Core/PowerPC/BreakPoints.cpp

@@ -11,8 +11,10 @@
 
 #include "Common/CommonTypes.h"
 #include "Common/DebugInterface.h"
+#include "Core/Core.h"
 #include "Core/PowerPC/JitCommon/JitBase.h"
 #include "Core/PowerPC/JitCommon/JitCache.h"
+#include "Core/PowerPC/PowerPC.h"
 
 bool BreakPoints::IsAddressBreakPoint(u32 address) const
 {
@@ -168,13 +170,18 @@ void MemChecks::AddFromStrings(const TMemChecksStr& mc_strings)
 
 void MemChecks::Add(const TMemCheck& memory_check)
 {
-  bool had_any = HasAny();
   if (GetMemCheck(memory_check.start_address) == nullptr)
+  {
+    bool had_any = HasAny();
     m_mem_checks.push_back(memory_check);
-  // If this is the first one, clear the JIT cache so it can switch to
-  // watchpoint-compatible code.
-  if (!had_any && g_jit)
-    g_jit->GetBlockCache()->SchedulateClearCacheThreadSafe();
+    bool lock = Core::PauseAndLock(true);
+    // If this is the first one, clear the JIT cache so it can switch to
+    // watchpoint-compatible code.
+    if (!had_any && g_jit)
+      g_jit->ClearCache();
+    PowerPC::DBATUpdated();
+    Core::PauseAndLock(false, lock);
+  }
 }
 
 void MemChecks::Remove(u32 address)
@@ -184,8 +191,11 @@ void MemChecks::Remove(u32 address)
     if (i->start_address == address)
     {
       m_mem_checks.erase(i);
+      bool lock = Core::PauseAndLock(true);
       if (!HasAny() && g_jit)
-        g_jit->GetBlockCache()->SchedulateClearCacheThreadSafe();
+        g_jit->ClearCache();
+      PowerPC::DBATUpdated();
+      Core::PauseAndLock(false, lock);
       return;
     }
   }

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

@@ -335,7 +335,7 @@ void EmuCodeBlock::MMIOLoadToReg(MMIO::Mapping* mmio, Gen::X64Reg reg_value,
 void EmuCodeBlock::SafeLoadToReg(X64Reg reg_value, const Gen::OpArg& opAddress, int accessSize,
                                  s32 offset, BitSet32 registersInUse, bool signExtend, int flags)
 {
-  bool slowmem = (flags & SAFE_LOADSTORE_FORCE_SLOWMEM) != 0 || g_jit->jo.alwaysUseMemFuncs;
+  bool slowmem = (flags & SAFE_LOADSTORE_FORCE_SLOWMEM) != 0;
 
   registersInUse[reg_value] = false;
   if (g_jit->jo.fastmem && !(flags & SAFE_LOADSTORE_NO_FASTMEM) && !slowmem)
@@ -492,7 +492,7 @@ void EmuCodeBlock::SafeWriteRegToReg(OpArg reg_value, X64Reg reg_addr, int acces
                                      BitSet32 registersInUse, int flags)
 {
   bool swap = !(flags & SAFE_LOADSTORE_NO_SWAP);
-  bool slowmem = (flags & SAFE_LOADSTORE_FORCE_SLOWMEM) != 0 || g_jit->jo.alwaysUseMemFuncs;
+  bool slowmem = (flags & SAFE_LOADSTORE_FORCE_SLOWMEM) != 0;
 
   // set the correct immediate format
   reg_value = FixImmediate(accessSize, reg_value);

Source/Core/Core/PowerPC/JitCommon/JitBase.cpp

@@ -46,7 +46,6 @@ bool JitBase::MergeAllowedNextInstructions(int count)
 void JitBase::UpdateMemoryOptions()
 {
   bool any_watchpoints = PowerPC::memchecks.HasAny();
-  jo.fastmem = SConfig::GetInstance().bFastmem && !any_watchpoints;
+  jo.fastmem = SConfig::GetInstance().bFastmem;
   jo.memcheck = SConfig::GetInstance().bMMU || any_watchpoints;
-  jo.alwaysUseMemFuncs = any_watchpoints;
 }

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

@@ -52,7 +52,6 @@ protected:
     bool accurateSinglePrecision;
     bool fastmem;
     bool memcheck;
-    bool alwaysUseMemFuncs;
   };
   struct JitState
   {

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

@@ -29,13 +29,6 @@
 
 using namespace Gen;
 
-static CoreTiming::EventType* s_clear_jit_cache_thread_safe;
-
-static void ClearCacheThreadSafe(u64 userdata, s64 cyclesdata)
-{
-  JitInterface::ClearCache();
-}
-
 bool JitBlock::OverlapsPhysicalRange(u32 address, u32 length) const
 {
   return physical_addresses.lower_bound(address) !=
@@ -50,7 +43,6 @@ JitBaseBlockCache::~JitBaseBlockCache() = default;
 
 void JitBaseBlockCache::Init()
 {
-  s_clear_jit_cache_thread_safe = CoreTiming::RegisterEvent("clearJitCache", ClearCacheThreadSafe);
   JitRegister::Init(SConfig::GetInstance().m_perfDir);
 
   Clear();
@@ -89,11 +81,6 @@ void JitBaseBlockCache::Reset()
   Init();
 }
 
-void JitBaseBlockCache::SchedulateClearCacheThreadSafe()
-{
-  CoreTiming::ScheduleEvent(0, s_clear_jit_cache_thread_safe, 0, CoreTiming::FromThread::NON_CPU);
-}
-
 JitBlock** JitBaseBlockCache::GetFastBlockMap()
 {
   return fast_block_map.data();

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

@@ -125,7 +125,6 @@ public:
   void Shutdown();
   void Clear();
   void Reset();
-  void SchedulateClearCacheThreadSafe();
 
   // Code Cache
   JitBlock** GetFastBlockMap();

Source/Core/Core/PowerPC/MMU.cpp

@@ -1132,6 +1132,24 @@ static TranslateAddressResult TranslatePageAddress(const u32 address, const XChe
   return TranslateAddressResult{TranslateAddressResult::PAGE_FAULT, 0};
 }
 
+static bool overlaps_memcheck(u32 pageEndAddress)
+{
+  if (!memchecks.HasAny())
+    return false;
+  u32 page_end_suffix = ((1 << BAT_INDEX_SHIFT)) - 1;
+  for (TMemCheck memcheck : memchecks.GetMemChecks())
+  {
+    if (((memcheck.start_address | page_end_suffix) == pageEndAddress ||
+         (memcheck.end_address | page_end_suffix) == pageEndAddress) ||
+        ((memcheck.start_address | page_end_suffix) < pageEndAddress &&
+         (memcheck.end_address | page_end_suffix) > pageEndAddress))
+    {
+      return true;
+    }
+  }
+  return false;
+}
+
 static void UpdateBATs(BatTable& bat_table, u32 base_spr)
 {
   // TODO: Separate BATs for MSR.PR==0 and MSR.PR==1
@@ -1191,6 +1209,8 @@ static void UpdateBATs(BatTable& bat_table, u32 base_spr)
           valid_bit = 0x3;
         else if ((address >> 28) == 0xE && (address < (0xE0000000 + Memory::L1_CACHE_SIZE)))
           valid_bit = 0x3;
+        if (overlaps_memcheck(((batu.BEPI | j) << BAT_INDEX_SHIFT) | ((1 << BAT_INDEX_SHIFT) - 1)))
+          valid_bit &= ~0x2;
 
         // (BEPI | j) == (BEPI & ~BL) | (j & BL).
         bat_table[batu.BEPI | j] = address | valid_bit;