JitArm64: Use EmitBackpatchRoutine more for psq_l/psq_st

In the case of the JitAsm routines, we can't actually use
backpatching. Still, I would like to gather all the load and
store instructions in one place to make future changes easier.

6 files changed, 269 insertions, 185 deletions

diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
index cf5712eaab..fa5ed3e67d 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
@@ -60,39 +60,22 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
   {
     if ((flags & BackPatchInfo::FLAG_STORE) && (flags & BackPatchInfo::FLAG_FLOAT))
     {
-      if ((flags & BackPatchInfo::FLAG_SIZE_32) && !(flags & BackPatchInfo::FLAG_PAIR))
-      {
-        m_float_emit.REV32(8, ARM64Reg::D0, RS);
-        m_float_emit.STR(32, ARM64Reg::D0, MEM_REG, addr);
-      }
-      else if ((flags & BackPatchInfo::FLAG_SIZE_32) && (flags & BackPatchInfo::FLAG_PAIR))
-      {
-        m_float_emit.REV32(8, ARM64Reg::D0, RS);
-        m_float_emit.STR(64, ARM64Reg::Q0, MEM_REG, addr);
-      }
-      else
-      {
-        m_float_emit.REV64(8, ARM64Reg::Q0, RS);
-        m_float_emit.STR(64, ARM64Reg::Q0, MEM_REG, addr);
-      }
+      ARM64Reg temp = ARM64Reg::D0;
+      temp = ByteswapBeforeStore(this, &m_float_emit, temp, EncodeRegToDouble(RS), flags, true);
+
+      m_float_emit.STR(BackPatchInfo::GetFlagSize(flags), temp, MEM_REG, addr);
     }
     else if ((flags & BackPatchInfo::FLAG_LOAD) && (flags & BackPatchInfo::FLAG_FLOAT))
     {
-      if (flags & BackPatchInfo::FLAG_SIZE_32)
-      {
-        m_float_emit.LDR(32, EncodeRegToDouble(RS), MEM_REG, addr);
-        m_float_emit.REV32(8, EncodeRegToDouble(RS), EncodeRegToDouble(RS));
-      }
-      else
-      {
-        m_float_emit.LDR(64, EncodeRegToDouble(RS), MEM_REG, addr);
-        m_float_emit.REV64(8, EncodeRegToDouble(RS), EncodeRegToDouble(RS));
-      }
+      m_float_emit.LDR(BackPatchInfo::GetFlagSize(flags), EncodeRegToDouble(RS), MEM_REG, addr);
+
+      ByteswapAfterLoad(this, &m_float_emit, EncodeRegToDouble(RS), EncodeRegToDouble(RS), flags,
+                        true, false);
     }
     else if (flags & BackPatchInfo::FLAG_STORE)
     {
       ARM64Reg temp = ARM64Reg::W0;
-      temp = ByteswapBeforeStore(this, temp, RS, flags, true);
+      temp = ByteswapBeforeStore(this, &m_float_emit, temp, RS, flags, true);
 
       if (flags & BackPatchInfo::FLAG_SIZE_32)
         STR(temp, MEM_REG, addr);
@@ -118,7 +101,7 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
       else if (flags & BackPatchInfo::FLAG_SIZE_8)
         LDRB(RS, MEM_REG, addr);
 
-      ByteswapAfterLoad(this, RS, RS, flags, true, false);
+      ByteswapAfterLoad(this, &m_float_emit, RS, RS, flags, true, false);
     }
   }
   const u8* fastmem_end = GetCodePtr();
@@ -158,52 +141,39 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
     ABI_PushRegisters(gprs_to_push);
     m_float_emit.ABI_PushRegisters(fprs_to_push, ARM64Reg::X30);
 
-    if ((flags & BackPatchInfo::FLAG_STORE) && (flags & BackPatchInfo::FLAG_FLOAT))
-    {
-      if ((flags & BackPatchInfo::FLAG_SIZE_32) && !(flags & BackPatchInfo::FLAG_PAIR))
-      {
-        m_float_emit.UMOV(32, ARM64Reg::W0, RS, 0);
-        MOVP2R(ARM64Reg::X8, &PowerPC::Write_U32);
-        BLR(ARM64Reg::X8);
-      }
-      else if ((flags & BackPatchInfo::FLAG_SIZE_32) && (flags & BackPatchInfo::FLAG_PAIR))
-      {
-        m_float_emit.UMOV(64, ARM64Reg::X0, RS, 0);
-        MOVP2R(ARM64Reg::X8, &PowerPC::Write_U64);
-        ROR(ARM64Reg::X0, ARM64Reg::X0, 32);
-        BLR(ARM64Reg::X8);
-      }
-      else
-      {
-        m_float_emit.UMOV(64, ARM64Reg::X0, RS, 0);
-        MOVP2R(ARM64Reg::X8, &PowerPC::Write_U64);
-        BLR(ARM64Reg::X8);
-      }
-    }
-    else if ((flags & BackPatchInfo::FLAG_LOAD) && (flags & BackPatchInfo::FLAG_FLOAT))
+    if (flags & BackPatchInfo::FLAG_STORE)
     {
-      if (flags & BackPatchInfo::FLAG_SIZE_32)
+      const u32 access_size = BackPatchInfo::GetFlagSize(flags);
+      ARM64Reg src_reg = RS;
+      const ARM64Reg dst_reg = access_size == 64 ? ARM64Reg::X0 : ARM64Reg::W0;
+
+      if (flags & BackPatchInfo::FLAG_FLOAT)
       {
-        MOVP2R(ARM64Reg::X8, &PowerPC::Read_U32);
-        BLR(ARM64Reg::X8);
-        m_float_emit.INS(32, RS, 0, ARM64Reg::X0);
+        if (access_size == 64)
+          m_float_emit.FMOV(dst_reg, EncodeRegToDouble(RS));
+        else
+          m_float_emit.FMOV(dst_reg, EncodeRegToSingle(RS));
+
+        src_reg = dst_reg;
       }
-      else
+
+      if (flags & BackPatchInfo::FLAG_PAIR)
       {
-        MOVP2R(ARM64Reg::X8, &PowerPC::Read_F64);
-        BLR(ARM64Reg::X8);
-        m_float_emit.INS(64, RS, 0, ARM64Reg::X0);
+        // Compensate for the Write_ functions swapping the whole write instead of each pair
+        SwapPairs(this, dst_reg, src_reg, flags);
+        src_reg = dst_reg;
       }
-    }
-    else if (flags & BackPatchInfo::FLAG_STORE)
-    {
-      MOV(ARM64Reg::W0, RS);
+
+      if (dst_reg != src_reg)
+        MOV(dst_reg, src_reg);
 
       const bool reverse = (flags & BackPatchInfo::FLAG_REVERSE) != 0;
 
-      if (flags & BackPatchInfo::FLAG_SIZE_32)
+      if (access_size == 64)
+        MOVP2R(ARM64Reg::X8, reverse ? &PowerPC::Write_U64_Swap : &PowerPC::Write_U64);
+      else if (access_size == 32)
         MOVP2R(ARM64Reg::X8, reverse ? &PowerPC::Write_U32_Swap : &PowerPC::Write_U32);
-      else if (flags & BackPatchInfo::FLAG_SIZE_16)
+      else if (access_size == 16)
         MOVP2R(ARM64Reg::X8, reverse ? &PowerPC::Write_U16_Swap : &PowerPC::Write_U16);
       else
         MOVP2R(ARM64Reg::X8, &PowerPC::Write_U8);
@@ -217,16 +187,40 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode, AR
     }
     else
     {
-      if (flags & BackPatchInfo::FLAG_SIZE_32)
+      const u32 access_size = BackPatchInfo::GetFlagSize(flags);
+
+      if (access_size == 64)
+        MOVP2R(ARM64Reg::X8, &PowerPC::Read_U64);
+      else if (access_size == 32)
         MOVP2R(ARM64Reg::X8, &PowerPC::Read_U32);
-      else if (flags & BackPatchInfo::FLAG_SIZE_16)
+      else if (access_size == 16)
         MOVP2R(ARM64Reg::X8, &PowerPC::Read_U16);
-      else if (flags & BackPatchInfo::FLAG_SIZE_8)
+      else
         MOVP2R(ARM64Reg::X8, &PowerPC::Read_U8);
 
       BLR(ARM64Reg::X8);
 
-      ByteswapAfterLoad(this, RS, ARM64Reg::W0, flags, false, false);
+      ARM64Reg src_reg = access_size == 64 ? ARM64Reg::X0 : ARM64Reg::W0;
+
+      if (flags & BackPatchInfo::FLAG_PAIR)
+      {
+        // Compensate for the Read_ functions swapping the whole read instead of each pair
+        const ARM64Reg dst_reg = flags & BackPatchInfo::FLAG_FLOAT ? src_reg : RS;
+        SwapPairs(this, dst_reg, src_reg, flags);
+        src_reg = dst_reg;
+      }
+
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+      {
+        if (access_size == 64)
+          m_float_emit.FMOV(EncodeRegToDouble(RS), src_reg);
+        else
+          m_float_emit.FMOV(EncodeRegToSingle(RS), src_reg);
+
+        src_reg = RS;
+      }
+
+      ByteswapAfterLoad(this, &m_float_emit, RS, src_reg, flags, false, false);
     }
 
     m_float_emit.ABI_PopRegisters(fprs_to_push, ARM64Reg::X30);
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp
index fc4603811f..0cedefac45 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStore.cpp
@@ -240,7 +240,7 @@ void JitArm64::SafeStoreFromReg(s32 dest, u32 value, s32 regOffset, u32 flags, s
     LDR(IndexType::Unsigned, ARM64Reg::X0, PPC_REG, PPCSTATE_OFF(gather_pipe_ptr));
 
     ARM64Reg temp = ARM64Reg::W1;
-    temp = ByteswapBeforeStore(this, temp, RS, flags, true);
+    temp = ByteswapBeforeStore(this, &m_float_emit, temp, RS, flags, true);
 
     if (accessSize == 32)
       STR(IndexType::Post, temp, ARM64Reg::X0, 4);
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
index da6af010d9..305eb8a9f2 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
@@ -42,7 +42,7 @@ void JitArm64::psq_lXX(UGeckoInstruction inst)
   constexpr ARM64Reg addr_reg = ARM64Reg::W0;
   constexpr ARM64Reg scale_reg = ARM64Reg::W1;
   constexpr ARM64Reg type_reg = ARM64Reg::W2;
-  ARM64Reg VS;
+  ARM64Reg VS = fpr.RW(inst.RS, RegType::Single);
 
   if (inst.RA || update)  // Always uses the register on update
   {
@@ -69,17 +69,20 @@ void JitArm64::psq_lXX(UGeckoInstruction inst)
 
   if (js.assumeNoPairedQuantize)
   {
-    VS = fpr.RW(inst.RS, RegType::Single);
+    BitSet32 gprs_in_use = gpr.GetCallerSavedUsed();
+    BitSet32 fprs_in_use = fpr.GetCallerSavedUsed();
+
+    // Wipe the registers we are using as temporaries
+    gprs_in_use &= BitSet32(~7);
+    fprs_in_use &= BitSet32(~3);
+    fprs_in_use[DecodeReg(VS)] = 0;
+
+    u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_32;
     if (!w)
-    {
-      ADD(EncodeRegTo64(addr_reg), EncodeRegTo64(addr_reg), MEM_REG);
-      m_float_emit.LD1(32, 1, EncodeRegToDouble(VS), EncodeRegTo64(addr_reg));
-    }
-    else
-    {
-      m_float_emit.LDR(32, VS, EncodeRegTo64(addr_reg), MEM_REG);
-    }
-    m_float_emit.REV32(8, EncodeRegToDouble(VS), EncodeRegToDouble(VS));
+      flags |= BackPatchInfo::FLAG_PAIR;
+
+    EmitBackpatchRoutine(flags, jo.fastmem, jo.fastmem, VS, EncodeRegTo64(addr_reg), gprs_in_use,
+                         fprs_in_use);
   }
   else
   {
@@ -91,7 +94,6 @@ void JitArm64::psq_lXX(UGeckoInstruction inst)
     LDR(EncodeRegTo64(type_reg), ARM64Reg::X30, ArithOption(EncodeRegTo64(type_reg), true));
     BLR(EncodeRegTo64(type_reg));
 
-    VS = fpr.RW(inst.RS, RegType::Single);
     m_float_emit.ORR(EncodeRegToDouble(VS), ARM64Reg::D0, ARM64Reg::D0);
   }
 
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
index 40129ce202..f083ba3dee 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitAsm.cpp
@@ -488,11 +488,14 @@ void JitArm64::GenerateQuantizedLoads()
 {
   // X0 is the address
   // X1 is the scale
+  // X2 is a temporary
   // X30 is LR
   // Q0 is the return
   // Q1 is a temporary
   ARM64Reg addr_reg = ARM64Reg::X0;
   ARM64Reg scale_reg = ARM64Reg::X1;
+  BitSet32 gprs_to_push = CALLER_SAVED_GPRS & ~BitSet32{0, 2};
+  BitSet32 fprs_to_push = BitSet32(0xFFFFFFFF) & ~BitSet32{0, 1};
   ARM64FloatEmitter float_emit(this);
 
   const u8* start = GetCodePtr();
@@ -500,15 +503,20 @@ void JitArm64::GenerateQuantizedLoads()
   BRK(100);
   const u8* loadPairedFloatTwo = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LD1(32, 1, ARM64Reg::D0, addr_reg);
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_32;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push & ~BitSet32{1},
+                         fprs_to_push);
     RET(ARM64Reg::X30);
   }
   const u8* loadPairedU8Two = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(16, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
+    constexpr u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_8;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.UXTL(8, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
@@ -521,8 +529,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedS8Two = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(16, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
+    constexpr u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_8;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.SXTL(8, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
@@ -535,9 +546,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedU16Two = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LD1(16, 1, ARM64Reg::D0, addr_reg);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_16;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.UXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
 
@@ -549,9 +562,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedS16Two = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LD1(16, 1, ARM64Reg::D0, addr_reg);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags = BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_16;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.SXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
 
@@ -564,15 +579,20 @@ void JitArm64::GenerateQuantizedLoads()
 
   const u8* loadPairedFloatOne = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(32, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_32;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push & ~BitSet32{1},
+                         fprs_to_push);
     RET(ARM64Reg::X30);
   }
   const u8* loadPairedU8One = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(8, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_8;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.UXTL(8, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
@@ -585,8 +605,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedS8One = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(8, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_8;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.SXTL(8, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
@@ -599,9 +622,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedU16One = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(16, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_16;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.UXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.UCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
 
@@ -613,9 +638,11 @@ void JitArm64::GenerateQuantizedLoads()
   }
   const u8* loadPairedS16One = GetCodePtr();
   {
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.LDR(16, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_LOAD | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_16;
+
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+
     float_emit.SXTL(16, ARM64Reg::D0, ARM64Reg::D0);
     float_emit.SCVTF(32, ARM64Reg::D0, ARM64Reg::D0);
 
@@ -663,6 +690,8 @@ void JitArm64::GenerateQuantizedStores()
   // Q1 is a temporary
   ARM64Reg scale_reg = ARM64Reg::X0;
   ARM64Reg addr_reg = ARM64Reg::X1;
+  BitSet32 gprs_to_push = CALLER_SAVED_GPRS & ~BitSet32{0, 1, 2};
+  BitSet32 fprs_to_push = BitSet32(0xFFFFFFFF) & ~BitSet32{0, 1};
   ARM64FloatEmitter float_emit(this);
 
   const u8* start = GetCodePtr();
@@ -671,17 +700,16 @@ void JitArm64::GenerateQuantizedStores()
   const u8* storePairedFloat;
   const u8* storePairedFloatSlow;
   {
+    constexpr u32 flags = BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_32;
+
     storePairedFloat = GetCodePtr();
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(64, ARM64Reg::Q0, 0, addr_reg, ARM64Reg::SP);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storePairedFloatSlow = GetCodePtr();
-    float_emit.UMOV(64, ARM64Reg::X0, ARM64Reg::Q0, 0);
-    ROR(ARM64Reg::X0, ARM64Reg::X0, 32);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U64);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
 
   const u8* storePairedU8;
@@ -698,18 +726,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.UQXTN(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags = BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_8;
+
     storePairedU8 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(16, ARM64Reg::Q0, 0, addr_reg, ARM64Reg::SP);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storePairedU8Slow = GetCodePtr();
     emit_quantize();
-    float_emit.UMOV(16, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    REV16(ARM64Reg::W0, ARM64Reg::W0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U16);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storePairedS8;
   const u8* storePairedS8Slow;
@@ -725,18 +753,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.SQXTN(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags = BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_8;
+
     storePairedS8 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(16, ARM64Reg::Q0, 0, addr_reg, ARM64Reg::SP);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storePairedS8Slow = GetCodePtr();
     emit_quantize();
-    float_emit.UMOV(16, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    REV16(ARM64Reg::W0, ARM64Reg::W0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U16);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
 
   const u8* storePairedU16;
@@ -750,21 +778,20 @@ void JitArm64::GenerateQuantizedStores()
 
       float_emit.FCVTZU(32, ARM64Reg::D0, ARM64Reg::D0);
       float_emit.UQXTN(16, ARM64Reg::D0, ARM64Reg::D0);
-      float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags = BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_16;
+
     storePairedU16 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(32, ARM64Reg::Q0, 0, addr_reg, ARM64Reg::SP);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storePairedU16Slow = GetCodePtr();
     emit_quantize();
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
-    float_emit.UMOV(32, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U32);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storePairedS16;  // Used by Viewtiful Joe's intro movie
   const u8* storePairedS16Slow;
@@ -777,36 +804,35 @@ void JitArm64::GenerateQuantizedStores()
 
       float_emit.FCVTZS(32, ARM64Reg::D0, ARM64Reg::D0);
       float_emit.SQXTN(16, ARM64Reg::D0, ARM64Reg::D0);
-      float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags = BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT |
+                          BackPatchInfo::FLAG_PAIR | BackPatchInfo::FLAG_SIZE_16;
+
     storePairedS16 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(32, ARM64Reg::Q0, 0, addr_reg, ARM64Reg::SP);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storePairedS16Slow = GetCodePtr();
     emit_quantize();
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
-    float_emit.UMOV(32, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U32);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
 
   const u8* storeSingleFloat;
   const u8* storeSingleFloatSlow;
   {
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_32;
+
     storeSingleFloat = GetCodePtr();
-    float_emit.REV32(8, ARM64Reg::D0, ARM64Reg::D0);
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.STR(32, IndexType::Unsigned, ARM64Reg::D0, addr_reg, 0);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storeSingleFloatSlow = GetCodePtr();
-    float_emit.UMOV(32, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U32);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storeSingleU8;  // Used by MKWii
   const u8* storeSingleU8Slow;
@@ -822,17 +848,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.UQXTN(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_8;
+
     storeSingleU8 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(8, ARM64Reg::Q0, 0, addr_reg);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storeSingleU8Slow = GetCodePtr();
     emit_quantize();
-    float_emit.UMOV(8, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U8);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storeSingleS8;
   const u8* storeSingleS8Slow;
@@ -848,17 +875,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.SQXTN(8, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_8;
+
     storeSingleS8 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.ST1(8, ARM64Reg::Q0, 0, addr_reg);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storeSingleS8Slow = GetCodePtr();
     emit_quantize();
-    float_emit.SMOV(8, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U8);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storeSingleU16;  // Used by MKWii
   const u8* storeSingleU16Slow;
@@ -873,18 +901,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.UQXTN(16, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_16;
+
     storeSingleU16 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
-    float_emit.ST1(16, ARM64Reg::Q0, 0, addr_reg);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storeSingleU16Slow = GetCodePtr();
     emit_quantize();
-    float_emit.UMOV(16, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U16);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
   const u8* storeSingleS16;
   const u8* storeSingleS16Slow;
@@ -899,18 +927,18 @@ void JitArm64::GenerateQuantizedStores()
       float_emit.SQXTN(16, ARM64Reg::D0, ARM64Reg::D0);
     };
 
+    constexpr u32 flags =
+        BackPatchInfo::FLAG_STORE | BackPatchInfo::FLAG_FLOAT | BackPatchInfo::FLAG_SIZE_16;
+
     storeSingleS16 = GetCodePtr();
     emit_quantize();
-    ADD(addr_reg, addr_reg, MEM_REG);
-    float_emit.REV16(8, ARM64Reg::D0, ARM64Reg::D0);
-    float_emit.ST1(16, ARM64Reg::Q0, 0, addr_reg);
+    EmitBackpatchRoutine(flags, true, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
     RET(ARM64Reg::X30);
 
     storeSingleS16Slow = GetCodePtr();
     emit_quantize();
-    float_emit.SMOV(16, ARM64Reg::W0, ARM64Reg::Q0, 0);
-    MOVP2R(ARM64Reg::X2, &PowerPC::Write_U16);
-    BR(ARM64Reg::X2);
+    EmitBackpatchRoutine(flags, false, false, ARM64Reg::D0, addr_reg, gprs_to_push, fprs_to_push);
+    RET(ARM64Reg::X30);
   }
 
   JitRegister::Register(start, GetCodePtr(), "JIT_QuantizedStore");
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
index 8c1237c061..f8aceb5615 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.cpp
@@ -191,19 +191,47 @@ private:
   bool m_sign_extend;
 };
 
-void ByteswapAfterLoad(ARM64XEmitter* emit, ARM64Reg dst_reg, ARM64Reg src_reg, u32 flags,
-                       bool is_reversed, bool is_extended)
+void SwapPairs(ARM64XEmitter* emit, ARM64Reg dst_reg, ARM64Reg src_reg, u32 flags)
+{
+  if (flags & BackPatchInfo::FLAG_SIZE_32)
+    emit->ROR(dst_reg, src_reg, 32);
+  else if (flags & BackPatchInfo::FLAG_SIZE_16)
+    emit->ROR(dst_reg, src_reg, 16);
+  else
+    emit->REV16(dst_reg, src_reg);
+}
+
+void ByteswapAfterLoad(ARM64XEmitter* emit, Arm64Gen::ARM64FloatEmitter* float_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)
+    if (flags & BackPatchInfo::FLAG_SIZE_64)
+    {
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV64(8, dst_reg, src_reg);
+      else
+        emit->REV64(dst_reg, src_reg);
+
+      src_reg = dst_reg;
+    }
+    else if (flags & BackPatchInfo::FLAG_SIZE_32)
     {
-      emit->REV32(dst_reg, src_reg);
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV32(8, dst_reg, src_reg);
+      else
+        emit->REV32(dst_reg, src_reg);
+
       src_reg = dst_reg;
     }
     else if (flags & BackPatchInfo::FLAG_SIZE_16)
     {
-      emit->REV16(dst_reg, src_reg);
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV16(8, dst_reg, src_reg);
+      else
+        emit->REV16(dst_reg, src_reg);
+
       src_reg = dst_reg;
     }
   }
@@ -215,25 +243,47 @@ void ByteswapAfterLoad(ARM64XEmitter* emit, ARM64Reg dst_reg, ARM64Reg src_reg,
   }
 
   if (dst_reg != src_reg)
-    emit->MOV(dst_reg, src_reg);
+  {
+    if (flags & BackPatchInfo::FLAG_FLOAT)
+      float_emit->ORR(dst_reg, src_reg, src_reg);
+    else
+      emit->MOV(dst_reg, src_reg);
+  }
 }
 
-ARM64Reg ByteswapBeforeStore(ARM64XEmitter* emit, ARM64Reg tmp_reg, ARM64Reg src_reg, u32 flags,
-                             bool want_reversed)
+ARM64Reg ByteswapBeforeStore(ARM64XEmitter* emit, Arm64Gen::ARM64FloatEmitter* float_emit,
+                             ARM64Reg tmp_reg, ARM64Reg src_reg, u32 flags, bool want_reversed)
 {
   ARM64Reg dst_reg = src_reg;
 
   if (want_reversed == !(flags & BackPatchInfo::FLAG_REVERSE))
   {
-    if (flags & BackPatchInfo::FLAG_SIZE_32)
+    if (flags & BackPatchInfo::FLAG_SIZE_64)
     {
       dst_reg = tmp_reg;
-      emit->REV32(dst_reg, src_reg);
+
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV64(8, dst_reg, src_reg);
+      else
+        emit->REV64(dst_reg, src_reg);
+    }
+    else if (flags & BackPatchInfo::FLAG_SIZE_32)
+    {
+      dst_reg = tmp_reg;
+
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV32(8, dst_reg, src_reg);
+      else
+        emit->REV32(dst_reg, src_reg);
     }
     else if (flags & BackPatchInfo::FLAG_SIZE_16)
     {
       dst_reg = tmp_reg;
-      emit->REV16(dst_reg, src_reg);
+
+      if (flags & BackPatchInfo::FLAG_FLOAT)
+        float_emit->REV16(8, dst_reg, src_reg);
+      else
+        emit->REV16(dst_reg, src_reg);
     }
   }
 
@@ -243,6 +293,8 @@ ARM64Reg ByteswapBeforeStore(ARM64XEmitter* emit, ARM64Reg tmp_reg, ARM64Reg src
 void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
                    BitSet32 fprs_in_use, ARM64Reg dst_reg, u32 address, u32 flags)
 {
+  ASSERT(!(flags & BackPatchInfo::FLAG_FLOAT));
+
   if (flags & BackPatchInfo::FLAG_SIZE_8)
   {
     MMIOReadCodeGenerator<u8> gen(emit, gprs_in_use, fprs_in_use, dst_reg, address,
@@ -262,13 +314,15 @@ void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32
     mmio->GetHandlerForRead<u32>(address).Visit(gen);
   }
 
-  ByteswapAfterLoad(emit, dst_reg, dst_reg, flags, false, true);
+  ByteswapAfterLoad(emit, nullptr, dst_reg, dst_reg, flags, false, true);
 }
 
 void MMIOWriteRegToAddr(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
                         BitSet32 fprs_in_use, ARM64Reg src_reg, u32 address, u32 flags)
 {
-  src_reg = ByteswapBeforeStore(emit, ARM64Reg::W1, src_reg, flags, false);
+  ASSERT(!(flags & BackPatchInfo::FLAG_FLOAT));
+
+  src_reg = ByteswapBeforeStore(emit, nullptr, ARM64Reg::W1, src_reg, flags, false);
 
   if (flags & BackPatchInfo::FLAG_SIZE_8)
   {
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
index 47d4b5ce95..9f1b8f8436 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit_Util.h
@@ -8,11 +8,17 @@
 
 #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 SwapPairs(Arm64Gen::ARM64XEmitter* emit, Arm64Gen::ARM64Reg dst_reg,
+               Arm64Gen::ARM64Reg src_reg, u32 flags);
 
-Arm64Gen::ARM64Reg ByteswapBeforeStore(Arm64Gen::ARM64XEmitter* emit, Arm64Gen::ARM64Reg tmp_reg,
-                                       Arm64Gen::ARM64Reg src_reg, u32 flags, bool want_reversed);
+void ByteswapAfterLoad(Arm64Gen::ARM64XEmitter* emit, Arm64Gen::ARM64FloatEmitter* float_emit,
+                       Arm64Gen::ARM64Reg dst_reg, Arm64Gen::ARM64Reg src_reg, u32 flags,
+                       bool is_reversed, bool is_extended);
+
+Arm64Gen::ARM64Reg ByteswapBeforeStore(Arm64Gen::ARM64XEmitter* emit,
+                                       Arm64Gen::ARM64FloatEmitter* float_emit,
+                                       Arm64Gen::ARM64Reg tmp_reg, Arm64Gen::ARM64Reg src_reg,
+                                       u32 flags, bool want_reversed);
 
 void MMIOLoadToReg(MMIO::Mapping* mmio, Arm64Gen::ARM64XEmitter* emit, BitSet32 gprs_in_use,
                    BitSet32 fprs_in_use, Arm64Gen::ARM64Reg dst_reg, u32 address, u32 flags);