From d3ef5d4ca4f14149c1f564d75348604de1b73e78 Mon Sep 17 00:00:00 2001
From: JosJuice <josjuice@gmail.com>
Date: Wed, 30 Jun 2021 23:04:24 +0200
Subject: [PATCH] JitArm64: Fix lwbrx/lhbrx with optimized MMIO access

---
 .../PowerPC/JitArm64/JitArm64_BackPatch.cpp   | 27 ++---------------
 .../Core/Core/PowerPC/JitArm64/Jit_Util.cpp   | 29 +++++++++++++++++++
 Source/Core/Core/PowerPC/JitArm64/Jit_Util.h  |  3 ++
 3 files changed, 35 insertions(+), 24 deletions(-)

diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
index 23236a524e..f2472a3eda 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
@@ -15,6 +15,7 @@
 
 #include "Core/HW/Memmap.h"
 #include "Core/PowerPC/JitArm64/Jit.h"
+#include "Core/PowerPC/JitArm64/Jit_Util.h"
 #include "Core/PowerPC/JitArmCommon/BackPatch.h"
 #include "Core/PowerPC/MMU.h"
 #include "Core/PowerPC/PowerPC.h"
@@ -120,16 +121,7 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
       else if (flags & BackPatchInfo::FLAG_SIZE_8)
         LDRB(RS, MEM_REG, addr);
 
-      if (!(flags & BackPatchInfo::FLAG_REVERSE))
-      {
-        if (flags & BackPatchInfo::FLAG_SIZE_32)
-          REV32(RS, RS);
-        else if (flags & BackPatchInfo::FLAG_SIZE_16)
-          REV16(RS, RS);
-      }
-
-      if (flags & BackPatchInfo::FLAG_EXTEND)
-        SXTH(RS, RS);
+      ByteswapAfterLoad(this, RS, RS, flags, true, false);
     }
   }
   const u8* fastmem_end = GetCodePtr();
@@ -235,20 +227,7 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
 
       BLR(ARM64Reg::X8);
 
-      if (!(flags & BackPatchInfo::FLAG_REVERSE))
-      {
-        MOV(RS, ARM64Reg::W0);
-      }
-      else
-      {
-        if (flags & BackPatchInfo::FLAG_SIZE_32)
-          REV32(RS, ARM64Reg::W0);
-        else if (flags & BackPatchInfo::FLAG_SIZE_16)
-          REV16(RS, ARM64Reg::W0);
-      }
-
-      if (flags & BackPatchInfo::FLAG_EXTEND)
-        SXTH(RS, RS);
+      ByteswapAfterLoad(this, RS, ARM64Reg::W0, flags, false, false);
     }
 
     m_float_emit.ABI_PopRegisters(fprs_to_push, ARM64Reg::X30);
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
index f04d87537c..33a4aa1409 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
@@ -192,6 +192,33 @@ private:
   bool m_sign_extend;
 };
 
+void ByteswapAfterLoad(ARM64XEmitter* emit, ARM64Reg dst_reg, ARM64Reg src_reg, u32 flags,
+                       bool is_reversed, bool is_extended)
+{
+  if (is_reversed == !(flags & BackPatchInfo::FLAG_REVERSE))
+  {
+    if (flags & BackPatchInfo::FLAG_SIZE_32)
+    {
+      emit->REV32(dst_reg, src_reg);
+      src_reg = dst_reg;
+    }
+    else if (flags & BackPatchInfo::FLAG_SIZE_16)
+    {
+      emit->REV16(dst_reg, src_reg);
+      src_reg = dst_reg;
+    }
+  }
+
+  if (!is_extended && (flags & BackPatchInfo::FLAG_EXTEND))
+  {
+    emit->SXTH(dst_reg, src_reg);
+    src_reg = dst_reg;
+  }
+
+  if (dst_reg != src_reg)
+    emit->MOV(dst_reg, src_reg);
+}
+
 void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
                    BitSet32 fprs_in_use, ARM64Reg dst_reg, u32 address, u32 flags)
 {
@@ -213,6 +240,8 @@ void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32
                                    flags & BackPatchInfo::FLAG_EXTEND);
     mmio->GetHandlerForRead<u32>(address).Visit(gen);
   }
+
+  ByteswapAfterLoad(emit, dst_reg, dst_reg, flags, false, true);
 }
 
 void MMIOWriteRegToAddr(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
index 98474d92bf..dc814d1353 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
@@ -9,6 +9,9 @@
 
 #include "Core/HW/MMIO.h"
 
+void ByteswapAfterLoad(Arm64Gen::ARM64XEmitter* emit, Arm64Gen::ARM64Reg dst_reg,
+                       Arm64Gen::ARM64Reg src_reg, u32 flags, bool is_reversed, bool is_extended);
+
 void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
                    BitSet32 fprs_in_use, Arm64Gen::ARM64Reg dst_reg, u32 address, u32 flags);