Jit64: Merge m_exception_handler_at_loc in TrampolineInfo

If MMU is enabled, there's an m_exception_handler_at_loc entry for almost every entry in m_back_patch_info, so keeping them separate is a huge waste: hashmaps aren't free.
Brings a decent improvement to codegen performance when MMU is enabled.

If MMU is disabled, however, this increases the size of TrampolineInfo by 8 bytes. It's a little sad, but it's important to prioritize the worst-case performance.
Also, it simplifies the next commit, which will reduce the memory usage anyway.

3 files changed, 6 insertions, 13 deletions

diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.cpp b/Source/Core/Core/PowerPC/Jit64/Jit.cpp
index 1248b1ab01..91085b6199 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.cpp
@@ -179,14 +179,6 @@ bool Jit64::BackPatch(SContext* ctx)
 
   TrampolineInfo& info = it->second;
 
-  u8* exceptionHandler = nullptr;
-  if (jo.memcheck)
-  {
-    auto it2 = m_exception_handler_at_loc.find(codePtr);
-    if (it2 != m_exception_handler_at_loc.end())
-      exceptionHandler = it2->second;
-  }
-
   // In the trampoline code, we jump back into the block at the beginning
   // of the next instruction. The next instruction comes immediately
   // after the backpatched operation, or BACKPATCH_SIZE bytes after the start
@@ -195,7 +187,7 @@ bool Jit64::BackPatch(SContext* ctx)
   // to insert the backpatch jump.)
 
   js.generatingTrampoline = true;
-  js.trampolineExceptionHandler = exceptionHandler;
+  js.trampolineExceptionHandler = info.exception_handler_at_loc;
   js.compilerPC = info.pc;
 
   // Generate the trampoline.
@@ -1203,12 +1195,11 @@ bool Jit64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
         SwitchToFarCode();
         if (!js.fastmemLoadStore)
         {
-          m_exception_handler_at_loc[js.fastmemLoadStore] = nullptr;
           SetJumpTarget(js.fixupExceptionHandler ? js.exceptionHandler : memException);
         }
         else
         {
-          m_exception_handler_at_loc[js.fastmemLoadStore] = GetWritableCodePtr();
+          m_back_patch_info.at(js.fastmemLoadStore).exception_handler_at_loc = GetWritableCodePtr();
         }
 
         RCForkGuard gpr_guard = gpr.Fork();
diff --git a/Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp b/Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp
index 1019aa7d0a..d99ec2b155 100644
--- a/Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp
+++ b/Source/Core/Core/PowerPC/Jit64Common/EmuCodeBlock.cpp
@@ -1061,5 +1061,4 @@ void EmuCodeBlock::SetFPRF(Gen::X64Reg xmm, bool single)
 void EmuCodeBlock::Clear()
 {
   m_back_patch_info.clear();
-  m_exception_handler_at_loc.clear();
 }
diff --git a/Source/Core/Core/PowerPC/Jit64Common/TrampolineInfo.h b/Source/Core/Core/PowerPC/Jit64Common/TrampolineInfo.h
index d76d3f724f..c230abe992 100644
--- a/Source/Core/Core/PowerPC/Jit64Common/TrampolineInfo.h
+++ b/Source/Core/Core/PowerPC/Jit64Common/TrampolineInfo.h
@@ -8,7 +8,7 @@
 #include "Common/x64Emitter.h"
 
 // Stores information we need to batch-patch a MOV with a call to the slow read/write path after
-// it faults. There will be 10s of thousands of these structs live, so be wary of making this too
+// it faults. There will be 100s of thousands of these structs live, so be wary of making this too
 // big.
 struct TrampolineInfo final
 {
@@ -46,4 +46,7 @@ struct TrampolineInfo final
   s32 offset = 0;
   Gen::X64Reg op_reg{};
   Gen::OpArg op_arg{};
+
+  // It's not set if memcheck is disabled. Otherwise, it's almost always set.
+  u8* exception_handler_at_loc = nullptr;
 };