JitArm64: Extract block exit link area generation.

Always use 3 instructions on linking.

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

@@ -325,20 +325,10 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
   linkData.call = LK;
   b->linkData.push_back(linkData);
 
-  MOVI2R(DISPATCHER_PC, destination);
+  blocks.WriteLinkBlock(*this, linkData);
 
-  if (!LK)
+  if (LK)
   {
-    B(dispatcher);
-  }
-  else
-  {
-    BL(dispatcher);
-
-    // MOVI2R might only require one instruction. So the const offset of 20 bytes
-    // might be wrong. Be sure and just add a NOP here.
-    HINT(HINT_NOP);
-
     // Write the regular exit node after the return.
     linkData.exitAddress = exit_address_after_return;
     linkData.exitPtrs = GetWritableCodePtr();
@@ -346,8 +336,7 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
     linkData.call = false;
     b->linkData.push_back(linkData);
 
-    MOVI2R(DISPATCHER_PC, exit_address_after_return);
+    blocks.WriteLinkBlock(*this, linkData);
-    B(dispatcher);
   }
 }
 
@@ -387,8 +376,7 @@ void JitArm64::WriteExit(Arm64Gen::ARM64Reg dest, bool LK, u32 exit_address_afte
     linkData.call = false;
     b->linkData.push_back(linkData);
 
-    MOVI2R(DISPATCHER_PC, exit_address_after_return);
+    blocks.WriteLinkBlock(*this, linkData);
-    B(dispatcher);
   }
 }
 
@@ -417,8 +405,7 @@ void JitArm64::FakeLKExit(u32 exit_address_after_return)
   linkData.call = false;
   b->linkData.push_back(linkData);
 
-  MOVI2R(DISPATCHER_PC, exit_address_after_return);
+  blocks.WriteLinkBlock(*this, linkData);
-  B(dispatcher);
 
   SetJumpTarget(skip_exit);
 }

Source/Core/Core/PowerPC/JitArm64/JitArm64Cache.cpp

@@ -2,9 +2,9 @@
 // Licensed under GPLv2+
 // Refer to the license.txt file included.
 
-#include "Core/PowerPC/JitArm64/Jit.h"
-#include "Common/CommonTypes.h"
 #include "Core/PowerPC/JitArm64/JitArm64Cache.h"
+#include "Common/CommonTypes.h"
+#include "Core/PowerPC/JitArm64/Jit.h"
 #include "Core/PowerPC/JitCommon/JitBase.h"
 #include "Core/PowerPC/JitInterface.h"
 
@@ -12,39 +12,57 @@ JitArm64BlockCache::JitArm64BlockCache(JitBase& jit) : JitBaseBlockCache{jit}
 {
 }
 
+void JitArm64BlockCache::WriteLinkBlock(Arm64Gen::ARM64XEmitter& emit,
+                                        const JitBlock::LinkData& source, const JitBlock* dest)
+{
+  if (!dest)
+  {
+    // Use a fixed amount of instructions, so we can assume to use 3 instructions on patching.
+    emit.MOVZ(DISPATCHER_PC, source.exitAddress & 0xFFFF, SHIFT_0);
+    emit.MOVK(DISPATCHER_PC, source.exitAddress >> 16, SHIFT_16);
+
+    if (source.call)
+      emit.BL(m_jit.GetAsmRoutines()->dispatcher);
+    else
+      emit.B(m_jit.GetAsmRoutines()->dispatcher);
+    return;
+  }
+
+  if (source.call)
+  {
+    // The "fast" BL must be the third instruction. So just use the former two to inline the
+    // downcount check here. It's better to do this near jump before the long jump to the other
+    // block.
+    FixupBranch fast_link = emit.B(CC_PL);
+    emit.BL(dest->checkedEntry);
+    emit.SetJumpTarget(fast_link);
+    emit.BL(dest->normalEntry);
+    return;
+  }
+
+  // Are we able to jump directly to the normal entry?
+  s64 distance = ((s64)dest->normalEntry - (s64)emit.GetCodePtr()) >> 2;
+  if (distance >= -0x40000 && distance <= 0x3FFFF)
+  {
+    emit.B(CC_PL, dest->normalEntry);
+    emit.B(dest->checkedEntry);
+    emit.BRK(101);
+    return;
+  }
+
+  FixupBranch fast_link = emit.B(CC_PL);
+  emit.B(dest->checkedEntry);
+  emit.SetJumpTarget(fast_link);
+  emit.B(dest->normalEntry);
+}
+
 void JitArm64BlockCache::WriteLinkBlock(const JitBlock::LinkData& source, const JitBlock* dest)
 {
   u8* location = source.exitPtrs;
   ARM64XEmitter emit(location);
 
-  if (dest)
+  WriteLinkBlock(emit, source, dest);
-  {
-    if (source.call)
-    {
-      emit.BL(dest->checkedEntry);
-    }
-    else
-    {
-      // Are we able to jump directly to the normal entry?
-      s64 distance = ((s64)dest->normalEntry - (s64)location) >> 2;
-      if (distance >= -0x40000 && distance <= 0x3FFFF)
-      {
-        emit.B(CC_PL, dest->normalEntry);
-      }
 
-      // Use the checked entry if either downcount is smaller zero,
-      // or if we're not able to inline the downcount check here.
-      emit.B(dest->checkedEntry);
-    }
-  }
-  else
-  {
-    emit.MOVI2R(DISPATCHER_PC, source.exitAddress);
-    if (source.call)
-      emit.BL(m_jit.GetAsmRoutines()->dispatcher);
-    else
-      emit.B(m_jit.GetAsmRoutines()->dispatcher);
-  }
   emit.FlushIcache();
 }
 

Source/Core/Core/PowerPC/JitArm64/JitArm64Cache.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include "Common/Arm64Emitter.h"
 #include "Core/PowerPC/JitCommon/JitCache.h"
 
 class JitBase;
@@ -15,6 +16,9 @@ class JitArm64BlockCache : public JitBaseBlockCache
 public:
   explicit JitArm64BlockCache(JitBase& jit);
 
+  void WriteLinkBlock(Arm64Gen::ARM64XEmitter& emit, const JitBlock::LinkData& source,
+                      const JitBlock* dest = nullptr);
+
 private:
   void WriteLinkBlock(const JitBlock::LinkData& source, const JitBlock* dest) override;
   void WriteDestroyBlock(const JitBlock& block) override;