Rebind textures if format changes or they're buffer textures

* Allocate work buffer for audio renderer instead of using guest supplied memory

* Typo

* Use GC.AllocateArray to allocate pinned array

.github/ISSUE_TEMPLATE/bug_report.md

@@ -1,6 +1,6 @@
 ---
 name: Bug Report
-about: Something doesn't work correctly in Ryujinx.
+about: Something doesn't work correctly in Ryujinx. Note that game-specific issues should be instead posted on the Game Compatibility List at https://github.com/Ryujinx/Ryujinx-Games-List, unless it is a provable regression.
 #assignees:
 ---
 

ARMeilleure/Common/AddressTable.cs

@@ -1,4 +1,4 @@
-using ARMeilleure.Diagnostics.EventSources;
+using ARMeilleure.Diagnostics;
 using System;
 using System.Collections.Generic;
 using System.Runtime.InteropServices;
@@ -218,7 +218,7 @@ namespace ARMeilleure.Common
 
             _pages.Add(page);
 
-            AddressTableEventSource.Log.Allocated(size, leaf);
+            TranslatorEventSource.Log.AddressTableAllocated(size, leaf);
 
             return page;
         }

ARMeilleure/Decoders/Decoder.cs

@@ -251,6 +251,13 @@ namespace ARMeilleure.Decoders
                 return false;
             }
 
+            // Compare and branch instructions are always conditional.
+            if (opCode.Instruction.Name == InstName.Cbz ||
+                opCode.Instruction.Name == InstName.Cbnz)
+            {
+                return false;
+            }
+
             // Note: On ARM32, most instructions have conditional execution,
             // so there's no "Always" (unconditional) branch like on ARM64.
             // We need to check if the condition is "Always" instead.

ARMeilleure/Decoders/IOpCode32Mem.cs

@@ -3,6 +3,7 @@ namespace ARMeilleure.Decoders
     interface IOpCode32Mem : IOpCode32
     {
         int Rt { get; }
+        int Rt2 => Rt | 1;
         int Rn { get; }
 
         bool WBack { get; }

ARMeilleure/Decoders/OpCode32SimdMovn.cs

@@ -0,0 +1,12 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCode32SimdMovn : OpCode32Simd
+    {
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovn(inst, address, opCode);
+
+        public OpCode32SimdMovn(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Size = (opCode >> 18) & 0x3;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16Adr.cs

@@ -4,7 +4,6 @@ namespace ARMeilleure.Decoders
     {
         public int Rd { get; }
 
-        public bool Add => true;
         public int Immediate { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16Adr(inst, address, opCode);

ARMeilleure/Decoders/OpCodeT32AluImm12.cs

@@ -0,0 +1,16 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluImm12 : OpCodeT32Alu, IOpCode32AluImm
+    {
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluImm12(inst, address, opCode);
+
+        public OpCodeT32AluImm12(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Immediate = (opCode & 0xff) | ((opCode >> 4) & 0x700) | ((opCode >> 15) & 0x800);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluReg.cs

@@ -0,0 +1,14 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluReg : OpCodeT32Alu, IOpCode32AluReg
+    {
+        public int Rm { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluReg(inst, address, opCode);
+
+        public OpCodeT32AluReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm = (opCode >> 0) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemImm8D.cs

@@ -23,7 +23,7 @@ namespace ARMeilleure.Decoders
             Add   = ((opCode >> 23) & 1) != 0;
             WBack = ((opCode >> 21) & 1) != 0;
 
-            Immediate = opCode & 0xff;
+            Immediate = (opCode & 0xff) << 2;
 
             IsLoad = ((opCode >> 20) & 1) != 0;
         }

ARMeilleure/Decoders/OpCodeT32MovImm16.cs

@@ -0,0 +1,19 @@
+using ARMeilleure.Common;
+using System.Runtime.Intrinsics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MovImm16 : OpCodeT32Alu, IOpCode32AluImm
+    {
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MovImm16(inst, address, opCode);
+
+        public OpCodeT32MovImm16(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Immediate = (opCode & 0xff) | ((opCode >> 4) & 0x700) | ((opCode >> 15) & 0x800) | ((opCode >> 4) & 0xf000);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32ShiftReg.cs

@@ -0,0 +1,19 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32ShiftReg : OpCodeT32Alu, IOpCode32AluRsReg
+    {
+        public int Rm => Rn;
+        public int Rs { get; }
+
+        public ShiftType ShiftType { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32ShiftReg(inst, address, opCode);
+
+        public OpCodeT32ShiftReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rs = (opCode >> 0) & 0xf;
+
+            ShiftType = (ShiftType)((opCode >> 21) & 3);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32Tb.cs

@@ -0,0 +1,16 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32Tb : OpCodeT32, IOpCode32BReg
+    {
+        public int Rm { get; }
+        public int Rn { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32Tb(inst, address, opCode);
+
+        public OpCodeT32Tb(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm = (opCode >> 0) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeTable.cs

@@ -809,8 +809,8 @@ namespace ARMeilleure.Decoders
             SetA32("111100100xxxxxxxxxxx1000xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_I,      OpCode32SimdReg.Create);
             SetA32("<<<<11100x11xxxxxxxx101xx0x0xxxx", InstName.Vadd,        InstEmit32.Vadd_S,      OpCode32SimdRegS.Create);
             SetA32("111100100x00xxxxxxxx1101xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0000x0x0xxxx", InstName.Vaddl,       InstEmit32.Vaddl_I,     OpCode32SimdRegLong.Create);
-            SetA32("1111001x1x<<xxxxxxxx0001x0x0xxxx", InstName.Vaddw,       InstEmit32.Vaddw_I,     OpCode32SimdRegWide.Create);
+            SetA32("1111001x1x<<xxxxxxx00000x0x0xxxx", InstName.Vaddl,       InstEmit32.Vaddl_I,     OpCode32SimdRegLong.Create);
+            SetA32("1111001x1x<<xxxxxxx00001x0x0xxxx", InstName.Vaddw,       InstEmit32.Vaddw_I,     OpCode32SimdRegWide.Create);
             SetA32("111100100x00xxxxxxxx0001xxx1xxxx", InstName.Vand,        InstEmit32.Vand_I,      OpCode32SimdBinary.Create);
             SetA32("111100100x01xxxxxxxx0001xxx1xxxx", InstName.Vbic,        InstEmit32.Vbic_I,      OpCode32SimdBinary.Create);
             SetA32("1111001x1x000xxxxxxx<<x10x11xxxx", InstName.Vbic,        InstEmit32.Vbic_II,     OpCode32SimdImm.Create);
@@ -879,6 +879,7 @@ namespace ARMeilleure.Decoders
             SetA32("111100100xxxxxxxxxxx1001xxx0xxxx", InstName.Vmla,        InstEmit32.Vmla_I,      OpCode32SimdReg.Create);
             SetA32("<<<<11100x00xxxxxxxx101xx0x0xxxx", InstName.Vmla,        InstEmit32.Vmla_S,      OpCode32SimdRegS.Create);
             SetA32("111100100x00xxxxxxxx1101xxx1xxxx", InstName.Vmla,        InstEmit32.Vmla_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxx01000x0x0xxxx", InstName.Vmlal,       InstEmit32.Vmlal_I,     OpCode32SimdRegLong.Create);
             SetA32("1111001x1x<<xxxxxxxx010xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_1,      OpCode32SimdRegElem.Create);
             SetA32("<<<<11100x00xxxxxxxx101xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_S,      OpCode32SimdRegS.Create);
             SetA32("111100100x10xxxxxxxx1101xxx1xxxx", InstName.Vmls,        InstEmit32.Vmls_V,      OpCode32SimdReg.Create);
@@ -898,7 +899,7 @@ namespace ARMeilleure.Decoders
             SetA32("1111001x1x001000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
             SetA32("1111001x1x010000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
             SetA32("1111001x1x100000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
-            SetA32("111100111x11xx10xxxx001000x0xxx0", InstName.Vmovn,       InstEmit32.Vmovn,       OpCode32SimdCmpZ.Create);
+            SetA32("111100111x11<<10xxxx001000x0xxx0", InstName.Vmovn,       InstEmit32.Vmovn,       OpCode32SimdMovn.Create);
             SetA32("<<<<11101111xxxxxxxx101000010000", InstName.Vmrs,        InstEmit32.Vmrs,        OpCode32SimdSpecial.Create);
             SetA32("<<<<11101110xxxxxxxx101000010000", InstName.Vmsr,        InstEmit32.Vmsr,        OpCode32SimdSpecial.Create);
             SetA32("1111001x1x<<xxxxxxxx100xx1x0xxxx", InstName.Vmul,        InstEmit32.Vmul_1,      OpCode32SimdRegElem.Create);
@@ -924,22 +925,33 @@ namespace ARMeilleure.Decoders
             SetA32("1111001x1x000xxxxxxx<<x10x01xxxx", InstName.Vorr,        InstEmit32.Vorr_II,     OpCode32SimdImm.Create);
             SetA32("111100100x<<xxxxxxxx1011x0x1xxxx", InstName.Vpadd,       InstEmit32.Vpadd_I,     OpCode32SimdReg.Create);
             SetA32("111100110x00xxxxxxxx1101x0x0xxxx", InstName.Vpadd,       InstEmit32.Vpadd_V,     OpCode32SimdReg.Create);
+            SetA32("111100111x11<<00xxxx0010xxx0xxxx", InstName.Vpaddl,      InstEmit32.Vpaddl,      OpCode32SimdCmpZ.Create);
             SetA32("1111001x0x<<xxxxxxxx1010x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_I,     OpCode32SimdReg.Create);
             SetA32("111100110x00xxxxxxxx1111x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_V,     OpCode32SimdReg.Create);
             SetA32("1111001x0x<<xxxxxxxx1010x0x1xxxx", InstName.Vpmin,       InstEmit32.Vpmin_I,     OpCode32SimdReg.Create);
             SetA32("111100110x10xxxxxxxx1111x0x0xxxx", InstName.Vpmin,       InstEmit32.Vpmin_V,     OpCode32SimdReg.Create);
+            SetA32("1111001x0xxxxxxxxxxx0000xxx1xxxx", InstName.Vqadd,       InstEmit32.Vqadd,       OpCode32SimdReg.Create);
+            SetA32("111100100x01xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create);
+            SetA32("111100100x10xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create);
+            SetA32("111100111x11<<10xxxx00101xx0xxx0", InstName.Vqmovn,      InstEmit32.Vqmovn,      OpCode32SimdMovn.Create);
+            SetA32("111100111x11<<10xxxx001001x0xxx0", InstName.Vqmovun,     InstEmit32.Vqmovun,     OpCode32SimdMovn.Create);
             SetA32("1111001x1x>>>xxxxxxx100101x1xxx0", InstName.Vqrshrn,     InstEmit32.Vqrshrn,     OpCode32SimdShImmNarrow.Create);
             SetA32("111100111x>>>xxxxxxx100001x1xxx0", InstName.Vqrshrun,    InstEmit32.Vqrshrun,    OpCode32SimdShImmNarrow.Create);
             SetA32("1111001x1x>>>xxxxxxx100100x1xxx0", InstName.Vqshrn,      InstEmit32.Vqshrn,      OpCode32SimdShImmNarrow.Create);
+            SetA32("111100111x>>>xxxxxxx100000x1xxx0", InstName.Vqshrun,     InstEmit32.Vqshrun,     OpCode32SimdShImmNarrow.Create);
+            SetA32("1111001x0xxxxxxxxxxx0010xxx1xxxx", InstName.Vqsub,       InstEmit32.Vqsub,       OpCode32SimdReg.Create);
             SetA32("111100111x111011xxxx010x0xx0xxxx", InstName.Vrecpe,      InstEmit32.Vrecpe,      OpCode32SimdSqrte.Create);
             SetA32("111100100x00xxxxxxxx1111xxx1xxxx", InstName.Vrecps,      InstEmit32.Vrecps,      OpCode32SimdReg.Create);
             SetA32("111100111x11xx00xxxx000<<xx0xxxx", InstName.Vrev,        InstEmit32.Vrev,        OpCode32SimdRev.Create);
+            SetA32("1111001x0x<<xxxxxxxx0001xxx0xxxx", InstName.Vrhadd,      InstEmit32.Vrhadd,      OpCode32SimdReg.Create);
             SetA32("111111101x1110xxxxxx101x01x0xxxx", InstName.Vrint,       InstEmit32.Vrint_RM,    OpCode32SimdS.Create);
             SetA32("<<<<11101x110110xxxx101x11x0xxxx", InstName.Vrint,       InstEmit32.Vrint_Z,     OpCode32SimdS.Create);
             SetA32("<<<<11101x110111xxxx101x01x0xxxx", InstName.Vrintx,      InstEmit32.Vrintx_S,    OpCode32SimdS.Create);
             SetA32("1111001x1x>>>xxxxxxx0010>xx1xxxx", InstName.Vrshr,       InstEmit32.Vrshr,       OpCode32SimdShImm.Create);
+            SetA32("111100101x>>>xxxxxxx100001x1xxx0", InstName.Vrshrn,      InstEmit32.Vrshrn,      OpCode32SimdShImmNarrow.Create);
             SetA32("111100111x111011xxxx010x1xx0xxxx", InstName.Vrsqrte,     InstEmit32.Vrsqrte,     OpCode32SimdSqrte.Create);
             SetA32("111100100x10xxxxxxxx1111xxx1xxxx", InstName.Vrsqrts,     InstEmit32.Vrsqrts,     OpCode32SimdReg.Create);
+            SetA32("1111001x1x>>>xxxxxxx0011>xx1xxxx", InstName.Vrsra,       InstEmit32.Vrsra,       OpCode32SimdShImm.Create);
             SetA32("111111100xxxxxxxxxxx101xx0x0xxxx", InstName.Vsel,        InstEmit32.Vsel,        OpCode32SimdSel.Create);
             SetA32("111100101x>>>xxxxxxx0101>xx1xxxx", InstName.Vshl,        InstEmit32.Vshl,        OpCode32SimdShImm.Create);
             SetA32("1111001x0xxxxxxxxxxx0100xxx0xxxx", InstName.Vshl,        InstEmit32.Vshl_I,      OpCode32SimdReg.Create);
@@ -970,7 +982,8 @@ namespace ARMeilleure.Decoders
             SetA32("111100110xxxxxxxxxxx1000xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_I,      OpCode32SimdReg.Create);
             SetA32("<<<<11100x11xxxxxxxx101xx1x0xxxx", InstName.Vsub,        InstEmit32.Vsub_S,      OpCode32SimdRegS.Create);
             SetA32("111100100x10xxxxxxxx1101xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0011x0x0xxxx", InstName.Vsubw,       InstEmit32.Vsubw_I,     OpCode32SimdRegWide.Create);
+            SetA32("1111001x1x<<xxxxxxx00010x0x0xxxx", InstName.Vsubl,       InstEmit32.Vsubl_I,     OpCode32SimdRegLong.Create);
+            SetA32("1111001x1x<<xxxxxxx00011x0x0xxxx", InstName.Vsubw,       InstEmit32.Vsubw_I,     OpCode32SimdRegWide.Create);
             SetA32("111100111x11xxxxxxxx10xxxxx0xxxx", InstName.Vtbl,        InstEmit32.Vtbl,        OpCode32SimdTbl.Create);
             SetA32("111100111x11<<10xxxx00001xx0xxxx", InstName.Vtrn,        InstEmit32.Vtrn,        OpCode32SimdCmpZ.Create);
             SetA32("111100100x<<xxxxxxxx1000xxx1xxxx", InstName.Vtst,        InstEmit32.Vtst,        OpCode32SimdReg.Create);
@@ -1008,7 +1021,7 @@ namespace ARMeilleure.Decoders
             SetT16("01000101xxxxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluRegHigh.Create);
             SetT16("01000110xxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16AluRegHigh.Create);
             SetT16("010001110xxxx000", InstName.Bx,     InstEmit32.Bx,      OpCodeT16BReg.Create);
-            SetT16("010001111xxxx000", InstName.Blx,    InstEmit32.Blx,     OpCodeT16BReg.Create);
+            SetT16("010001111xxxx000", InstName.Blx,    InstEmit32.Blxr,    OpCodeT16BReg.Create);
             SetT16("01001xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemLit.Create);
             SetT16("0101000xxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemReg.Create);
             SetT16("0101001xxxxxxxxx", InstName.Strh,   InstEmit32.Strh,    OpCodeT16MemReg.Create);
@@ -1056,6 +1069,7 @@ namespace ARMeilleure.Decoders
             SetT32("11110x01010xxxxx0xxxxxxxxxxxxxxx", InstName.Adc,      InstEmit32.Adc,      OpCodeT32AluImm.Create);
             SetT32("11101011000<xxxx0xxx<<<<xxxxxxxx", InstName.Add,      InstEmit32.Add,      OpCodeT32AluRsImm.Create);
             SetT32("11110x01000<xxxx0xxx<<<<xxxxxxxx", InstName.Add,      InstEmit32.Add,      OpCodeT32AluImm.Create);
+            SetT32("11110x100000xxxx0xxxxxxxxxxxxxxx", InstName.Add,      InstEmit32.Add,      OpCodeT32AluImm12.Create);
             SetT32("11101010000<xxxx0xxx<<<<xxxxxxxx", InstName.And,      InstEmit32.And,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00000<xxxx0xxx<<<<xxxxxxxx", InstName.And,      InstEmit32.And,      OpCodeT32AluImm.Create);
             SetT32("11110x<<<xxxxxxx10x0xxxxxxxxxxxx", InstName.B,        InstEmit32.B,        OpCodeT32BImm20.Create);
@@ -1064,12 +1078,14 @@ namespace ARMeilleure.Decoders
             SetT32("11110x00001xxxxx0xxxxxxxxxxxxxxx", InstName.Bic,      InstEmit32.Bic,      OpCodeT32AluImm.Create);
             SetT32("11110xxxxxxxxxxx11x1xxxxxxxxxxxx", InstName.Bl,       InstEmit32.Bl,       OpCodeT32BImm24.Create);
             SetT32("11110xxxxxxxxxxx11x0xxxxxxxxxxx0", InstName.Blx,      InstEmit32.Blx,      OpCodeT32BImm24.Create);
+            SetT32("111110101011xxxx1111xxxx1000xxxx", InstName.Clz,      InstEmit32.Clz,      OpCodeT32AluReg.Create);
             SetT32("111010110001xxxx0xxx1111xxxxxxxx", InstName.Cmn,      InstEmit32.Cmn,      OpCodeT32AluRsImm.Create);
             SetT32("11110x010001xxxx0xxx1111xxxxxxxx", InstName.Cmn,      InstEmit32.Cmn,      OpCodeT32AluImm.Create);
             SetT32("111010111011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluRsImm.Create);
             SetT32("11110x011011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluImm.Create);
             SetT32("11101010100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluImm.Create);
+            SetT32("111010001101xxxxxxxx111110101111", InstName.Lda,      InstEmit32.Lda,      OpCodeT32MemLdEx.Create);
             SetT32("111010001101xxxxxxxx111111101111", InstName.Ldaex,    InstEmit32.Ldaex,    OpCodeT32MemLdEx.Create);
             SetT32("1110100010x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
             SetT32("1110100100x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
@@ -1082,7 +1098,8 @@ namespace ARMeilleure.Decoders
             SetT32("111110000001xxxx<<<<1100xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
             SetT32("111110000001xxxx<<<<11x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
             SetT32("111110001001xxxxxxxxxxxxxxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm12.Create);
-            SetT32("1110100>x1>1<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
+            SetT32("11101000x111<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
+            SetT32("11101001x1x1<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
             SetT32("111110000011xxxx<<<<10x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
             SetT32("111110000011xxxx<<<<1100xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
             SetT32("111110000011xxxx<<<<11x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
@@ -1096,9 +1113,12 @@ namespace ARMeilleure.Decoders
             SetT32("111110010011xxxx<<<<11x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
             SetT32("111110011011xxxxxxxxxxxxxxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm12.Create);
             SetT32("11101010010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluRsImm.Create);
+            SetT32("111110100xxxxxxx1111xxxx0000xxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32ShiftReg.Create);
             SetT32("11110x00010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluImm.Create);
+            SetT32("11110x100100xxxx0xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32MovImm16.Create);
             SetT32("11101010011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluImm.Create);
+            SetT32("11110011101011111000000000000000", InstName.Nop,      InstEmit32.Nop,      OpCodeT32.Create);
             SetT32("11101010011x<<<<0xxxxxxxxxxxxxxx", InstName.Orn,      InstEmit32.Orn,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00011x<<<<0xxxxxxxxxxxxxxx", InstName.Orn,      InstEmit32.Orn,      OpCodeT32AluImm.Create);
             SetT32("11101010010x<<<<0xxxxxxxxxxxxxxx", InstName.Orr,      InstEmit32.Orr,      OpCodeT32AluRsImm.Create);
@@ -1115,11 +1135,14 @@ namespace ARMeilleure.Decoders
             SetT32("111110000100<<<<xxxx000000xxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemRsImm.Create);
             SetT32("111110000000xxxxxxxx1<<>xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
             SetT32("111110001000xxxxxxxxxxxxxxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm12.Create);
-            SetT32("1110100>x1>0<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
+            SetT32("11101000x110<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
+            SetT32("11101001x1x0<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
             SetT32("111110000010xxxxxxxx1<<>xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
             SetT32("111110001010xxxxxxxxxxxxxxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm12.Create);
             SetT32("11101011101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluRsImm.Create);
             SetT32("11110x01101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluImm.Create);
+            SetT32("111010001101xxxx111100000000xxxx", InstName.Tbb,      InstEmit32.Tbb,      OpCodeT32Tb.Create);
+            SetT32("111010001101xxxx111100000001xxxx", InstName.Tbh,      InstEmit32.Tbh,      OpCodeT32Tb.Create);
             SetT32("111010101001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluRsImm.Create);
             SetT32("11110x001001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluImm.Create);
             SetT32("111010100001xxxx0xxx1111xxxxxxxx", InstName.Tst,      InstEmit32.Tst,      OpCodeT32AluRsImm.Create);

ARMeilleure/Diagnostics/EventSources/AddressTableEventSource.cs

@@ -1,51 +0,0 @@
-using System.Diagnostics.Tracing;
-
-namespace ARMeilleure.Diagnostics.EventSources
-{
-    [EventSource(Name = "ARMeilleure")]
-    class AddressTableEventSource : EventSource
-    {
-        public static readonly AddressTableEventSource Log = new();
-
-        private ulong _size;
-        private ulong _leafSize;
-        private PollingCounter _sizeCounter;
-        private PollingCounter _leafSizeCounter;
-
-        public AddressTableEventSource()
-        {
-            _sizeCounter = new PollingCounter("addr-tab-alloc", this, () => _size / 1024d / 1024d)
-            {
-                DisplayName = "AddressTable Total Bytes Allocated",
-                DisplayUnits = "MB"
-            };
-
-            _leafSizeCounter = new PollingCounter("addr-tab-leaf-alloc", this, () => _leafSize / 1024d / 1024d)
-            {
-                DisplayName = "AddressTable Total Leaf Bytes Allocated",
-                DisplayUnits = "MB"
-            };
-        }
-
-        public void Allocated(int bytes, bool leaf)
-        {
-            _size += (uint)bytes;
-
-            if (leaf)
-            {
-                _leafSize += (uint)bytes;
-            }
-        }
-
-        protected override void Dispose(bool disposing)
-        {
-            _leafSizeCounter.Dispose();
-            _leafSizeCounter = null;
-
-            _sizeCounter.Dispose();
-            _sizeCounter = null;
-
-            base.Dispose(disposing);
-        }
-    }
-}

ARMeilleure/Diagnostics/TranslatorEventSource.cs

@@ -0,0 +1,67 @@
+using System.Diagnostics.Tracing;
+using System.Threading;
+
+namespace ARMeilleure.Diagnostics
+{
+    [EventSource(Name = "ARMeilleure")]
+    class TranslatorEventSource : EventSource
+    {
+        public static readonly TranslatorEventSource Log = new();
+
+        private int _rejitQueue;
+        private ulong _funcTabSize;
+        private ulong _funcTabLeafSize;
+        private PollingCounter _rejitQueueCounter;
+        private PollingCounter _funcTabSizeCounter;
+        private PollingCounter _funcTabLeafSizeCounter;
+
+        public TranslatorEventSource()
+        {
+            _rejitQueueCounter = new PollingCounter("rejit-queue-length", this, () => _rejitQueue)
+            {
+                DisplayName = "Rejit Queue Length"
+            };
+
+            _funcTabSizeCounter = new PollingCounter("addr-tab-alloc", this, () => _funcTabSize / 1024d / 1024d)
+            {
+                DisplayName = "AddressTable Total Bytes Allocated",
+                DisplayUnits = "MB"
+            };
+
+            _funcTabLeafSizeCounter = new PollingCounter("addr-tab-leaf-alloc", this, () => _funcTabLeafSize / 1024d / 1024d)
+            {
+                DisplayName = "AddressTable Total Leaf Bytes Allocated",
+                DisplayUnits = "MB"
+            };
+        }
+
+        public void RejitQueueAdd(int count)
+        {
+            Interlocked.Add(ref _rejitQueue, count);
+        }
+
+        public void AddressTableAllocated(int bytes, bool leaf)
+        {
+            _funcTabSize += (uint)bytes;
+
+            if (leaf)
+            {
+                _funcTabLeafSize += (uint)bytes;
+            }
+        }
+
+        protected override void Dispose(bool disposing)
+        {
+            _rejitQueueCounter.Dispose();
+            _rejitQueueCounter = null;
+
+            _funcTabLeafSizeCounter.Dispose();
+            _funcTabLeafSizeCounter = null;
+
+            _funcTabSizeCounter.Dispose();
+            _funcTabSizeCounter = null;
+
+            base.Dispose(disposing);
+        }
+    }
+}

ARMeilleure/Instructions/InstEmitFlow32.cs

@@ -107,5 +107,30 @@ namespace ARMeilleure.Instructions
 
             context.SetIfThenBlockState(op.IfThenBlockConds);
         }
+
+        public static void Tbb(ArmEmitterContext context) => EmitTb(context, halfword: false);
+        public static void Tbh(ArmEmitterContext context)  => EmitTb(context, halfword: true);
+
+        private static void EmitTb(ArmEmitterContext context, bool halfword)
+        {
+            OpCodeT32Tb op = (OpCodeT32Tb)context.CurrOp;
+
+            Operand halfwords;
+
+            if (halfword)
+            {
+                Operand address = context.Add(GetIntA32(context, op.Rn), context.ShiftLeft(GetIntA32(context, op.Rm), Const(1)));
+                halfwords = InstEmitMemoryHelper.EmitReadInt(context, address, 1);
+            }
+            else
+            {
+                Operand address = context.Add(GetIntA32(context, op.Rn), GetIntA32(context, op.Rm));
+                halfwords = InstEmitMemoryHelper.EmitReadIntAligned(context, address, 0);
+            }
+
+            Operand targetAddress = context.Add(Const((int)op.GetPc()), context.ShiftLeft(halfwords, Const(1)));
+
+            EmitVirtualJump(context, targetAddress, isReturn: false);
+        }
     }
 }

ARMeilleure/Instructions/InstEmitMemory32.cs

@@ -205,13 +205,13 @@ namespace ARMeilleure.Instructions
                     context.BranchIfTrue(lblBigEndian, GetFlag(PState.EFlag));
 
                     Load(op.Rt, 0, WordSizeLog2);
-                    Load(op.Rt | 1, 4, WordSizeLog2);
+                    Load(op.Rt2, 4, WordSizeLog2);
 
                     context.Branch(lblEnd);
 
                     context.MarkLabel(lblBigEndian);
 
-                    Load(op.Rt | 1, 0, WordSizeLog2);
+                    Load(op.Rt2, 0, WordSizeLog2);
                     Load(op.Rt, 4, WordSizeLog2);
 
                     context.MarkLabel(lblEnd);
@@ -238,13 +238,13 @@ namespace ARMeilleure.Instructions
                     context.BranchIfTrue(lblBigEndian, GetFlag(PState.EFlag));
 
                     Store(op.Rt, 0, WordSizeLog2);
-                    Store(op.Rt | 1, 4, WordSizeLog2);
+                    Store(op.Rt2, 4, WordSizeLog2);
 
                     context.Branch(lblEnd);
 
                     context.MarkLabel(lblBigEndian);
 
-                    Store(op.Rt | 1, 0, WordSizeLog2);
+                    Store(op.Rt2, 0, WordSizeLog2);
                     Store(op.Rt, 4, WordSizeLog2);
 
                     context.MarkLabel(lblEnd);

ARMeilleure/Instructions/InstEmitMemoryHelper.cs

@@ -123,6 +123,41 @@ namespace ARMeilleure.Instructions
                    context.CurrOp is OpCodeSimdMemSs);
         }
 
+        public static Operand EmitReadInt(ArmEmitterContext context, Operand address, int size)
+        {
+            Operand temp = context.AllocateLocal(size == 3 ? OperandType.I64 : OperandType.I32);
+
+            Operand lblSlowPath = Label();
+            Operand lblEnd      = Label();
+
+            Operand physAddr = EmitPtPointerLoad(context, address, lblSlowPath, write: false, size);
+
+            Operand value = default;
+
+            switch (size)
+            {
+                case 0: value = context.Load8 (physAddr);                  break;
+                case 1: value = context.Load16(physAddr);                  break;
+                case 2: value = context.Load  (OperandType.I32, physAddr); break;
+                case 3: value = context.Load  (OperandType.I64, physAddr); break;
+            }
+
+            context.Copy(temp, value);
+
+            if (!context.Memory.Type.IsHostMapped())
+            {
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblSlowPath, BasicBlockFrequency.Cold);
+
+                context.Copy(temp, EmitReadIntFallback(context, address, size));
+
+                context.MarkLabel(lblEnd);
+            }
+
+            return temp;
+        }
+
         private static void EmitReadInt(ArmEmitterContext context, Operand address, int rt, int size)
         {
             Operand lblSlowPath = Label();
@@ -419,6 +454,11 @@ namespace ARMeilleure.Instructions
         }
 
         private static void EmitReadIntFallback(ArmEmitterContext context, Operand address, int rt, int size)
+        {
+            SetInt(context, rt, EmitReadIntFallback(context, address, size));
+        }
+
+        private static Operand EmitReadIntFallback(ArmEmitterContext context, Operand address, int size)
         {
             MethodInfo info = null;
 
@@ -430,7 +470,7 @@ namespace ARMeilleure.Instructions
                 case 3: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadUInt64)); break;
             }
 
-            SetInt(context, rt, context.Call(info, address));
+            return context.Call(info, address);
         }
 
         private static void EmitReadVectorFallback(

ARMeilleure/Instructions/InstEmitSimdArithmetic32.cs

@@ -777,6 +777,13 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void Vmlal_I(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            EmitVectorTernaryLongOpI32(context, (d, n, m) => context.Add(d, context.Multiply(n, m)), !op.U);
+        }
+
         public static void Vmls_S(ArmEmitterContext context)
         {
             if (Optimizations.FastFP && Optimizations.UseSse2)
@@ -992,6 +999,13 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void Vpaddl(ArmEmitterContext context)
+        {
+            OpCode32Simd op = (OpCode32Simd)context.CurrOp;
+
+            EmitVectorPairwiseLongOpI32(context, (op1, op2) => context.Add(op1, op2), (op.Opc & 1) == 0);
+        }
+
         public static void Vpmax_V(ArmEmitterContext context)
         {
             if (Optimizations.FastFP && Optimizations.UseSse2)
@@ -1052,6 +1066,62 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void Vqadd(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            EmitSaturatingAddSubBinaryOp(context, add: true, !op.U);
+        }
+
+        public static void Vqdmulh(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+            int eSize = 8 << op.Size;
+
+            EmitVectorBinaryOpI32(context, (op1, op2) =>
+            {
+                if (op.Size == 2)
+                {
+                    op1 = context.SignExtend32(OperandType.I64, op1);
+                    op2 = context.SignExtend32(OperandType.I64, op2);
+                }
+
+                Operand res = context.Multiply(op1, op2);
+                res = context.ShiftRightSI(res, Const(eSize - 1));
+                res = EmitSatQ(context, res, eSize, signedSrc: true, signedDst: true);
+
+                if (op.Size == 2)
+                {
+                    res = context.ConvertI64ToI32(res);
+                }
+
+                return res;
+            }, signed: true);
+        }
+
+        public static void Vqmovn(ArmEmitterContext context)
+        {
+            OpCode32SimdMovn op = (OpCode32SimdMovn)context.CurrOp;
+
+            bool signed = !op.Q;
+
+            EmitVectorUnaryNarrowOp32(context, (op1) => EmitSatQ(context, op1, 8 << op.Size, signed, signed), signed);
+        }
+
+        public static void Vqmovun(ArmEmitterContext context)
+        {
+            OpCode32SimdMovn op = (OpCode32SimdMovn)context.CurrOp;
+
+            EmitVectorUnaryNarrowOp32(context, (op1) => EmitSatQ(context, op1, 8 << op.Size, signedSrc: true, signedDst: false), signed: true);
+        }
+
+        public static void Vqsub(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            EmitSaturatingAddSubBinaryOp(context, add: false, !op.U);
+        }
+
         public static void Vrev(ArmEmitterContext context)
         {
             OpCode32SimdRev op = (OpCode32SimdRev)context.CurrOp;
@@ -1202,6 +1272,30 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void Vrhadd(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            EmitVectorBinaryOpI32(context, (op1, op2) =>
+            {
+                if (op.Size == 2)
+                {
+                    op1 = context.ZeroExtend32(OperandType.I64, op1);
+                    op2 = context.ZeroExtend32(OperandType.I64, op2);
+                }
+
+                Operand res = context.Add(context.Add(op1, op2), Const(op1.Type, 1L));
+                res = context.ShiftRightUI(res, Const(1));
+
+                if (op.Size == 2)
+                {
+                    res = context.ConvertI64ToI32(res);
+                }
+
+                return res;
+            }, !op.U);
+        }
+
         public static void Vrsqrte(ArmEmitterContext context)
         {
             OpCode32SimdSqrte op = (OpCode32SimdSqrte)context.CurrOp;
@@ -1349,6 +1443,13 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void Vsubl_I(ArmEmitterContext context)
+        {
+            OpCode32SimdRegLong op = (OpCode32SimdRegLong)context.CurrOp;
+
+            EmitVectorBinaryLongOpI32(context, (op1, op2) => context.Subtract(op1, op2), !op.U);
+        }
+
         public static void Vsubw_I(ArmEmitterContext context)
         {
             OpCode32SimdRegWide op = (OpCode32SimdRegWide)context.CurrOp;
@@ -1356,6 +1457,46 @@ namespace ARMeilleure.Instructions
             EmitVectorBinaryWideOpI32(context, (op1, op2) => context.Subtract(op1, op2), !op.U);
         }
 
+        private static void EmitSaturatingAddSubBinaryOp(ArmEmitterContext context, bool add, bool signed)
+        {
+            OpCode32Simd op = (OpCode32Simd)context.CurrOp;
+
+            EmitVectorBinaryOpI32(context, (ne, me) =>
+            {
+                if (op.Size <= 2)
+                {
+                    if (op.Size == 2)
+                    {
+                        ne = signed ? context.SignExtend32(OperandType.I64, ne) : context.ZeroExtend32(OperandType.I64, ne);
+                        me = signed ? context.SignExtend32(OperandType.I64, me) : context.ZeroExtend32(OperandType.I64, me);
+                    }
+
+                    Operand res = add ? context.Add(ne, me) : context.Subtract(ne, me);
+
+                    res = EmitSatQ(context, res, 8 << op.Size, signedSrc: true, signed);
+
+                    if (op.Size == 2)
+                    {
+                        res = context.ConvertI64ToI32(res);
+                    }
+
+                    return res;
+                }
+                else if (add) /* if (op.Size == 3) */
+                {
+                    return signed
+                        ? EmitBinarySignedSatQAdd(context, ne, me)
+                        : EmitBinaryUnsignedSatQAdd(context, ne, me);
+                }
+                else /* if (sub) */
+                {
+                    return signed
+                        ? EmitBinarySignedSatQSub(context, ne, me)
+                        : EmitBinaryUnsignedSatQSub(context, ne, me);
+                }
+            }, signed);
+        }
+
         private static void EmitSse41MaxMinNumOpF32(ArmEmitterContext context, bool isMaxNum, bool scalar)
         {
             IOpCode32Simd op = (IOpCode32Simd)context.CurrOp;

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -1352,7 +1352,7 @@ namespace ARMeilleure.Instructions
 
                 if (op.Size <= 2)
                 {
-                    de = EmitSatQ(context, emit(ne), op.Size, signedSrc: true, signedDst: true);
+                    de = EmitSignedSrcSatQ(context, emit(ne), op.Size, signedDst: true);
                 }
                 else /* if (op.Size == 3) */
                 {
@@ -1419,15 +1419,18 @@ namespace ARMeilleure.Instructions
                     {
                         Operand temp = add ? context.Add(ne, me) : context.Subtract(ne, me);
 
-                        de = EmitSatQ(context, temp, op.Size, signedSrc: true, signedDst: signed);
+                        de = EmitSignedSrcSatQ(context, temp, op.Size, signedDst: signed);
                     }
-                    else if (add) /* if (op.Size == 3) */
+                    else /* if (op.Size == 3) */
                     {
-                        de = EmitBinarySatQAdd(context, ne, me, signed);
+                        if (add)
+                        {
+                            de = signed ? EmitBinarySignedSatQAdd(context, ne, me) : EmitBinaryUnsignedSatQAdd(context, ne, me);
                         }
                         else /* if (sub) */
                         {
-                        de = EmitBinarySatQSub(context, ne, me, signed);
+                            de = signed ? EmitBinarySignedSatQSub(context, ne, me) : EmitBinaryUnsignedSatQSub(context, ne, me);
+                        }
                     }
 
                     res = EmitVectorInsert(context, res, de, index, op.Size);
@@ -1445,11 +1448,11 @@ namespace ARMeilleure.Instructions
                     {
                         Operand temp = context.Add(ne, me);
 
-                        de = EmitSatQ(context, temp, op.Size, signedSrc: true, signedDst: signed);
+                        de = EmitSignedSrcSatQ(context, temp, op.Size, signedDst: signed);
                     }
                     else /* if (op.Size == 3) */
                     {
-                        de = EmitBinarySatQAccumulate(context, ne, me, signed);
+                        de = signed ? EmitBinarySignedSatQAcc(context, ne, me) : EmitBinaryUnsignedSatQAcc(context, ne, me);
                     }
 
                     res = EmitVectorInsert(context, res, de, index, op.Size);
@@ -1475,7 +1478,7 @@ namespace ARMeilleure.Instructions
                         me = EmitVectorExtract(context, ((OpCodeSimdReg)op).Rm, index, op.Size, signed);
                     }
 
-                    Operand de = EmitSatQ(context, emit(ne, me), op.Size, true, signed);
+                    Operand de = EmitSignedSrcSatQ(context, emit(ne, me), op.Size, signedDst: signed);
 
                     res = EmitVectorInsert(context, res, de, index, op.Size);
                 }
@@ -1520,7 +1523,9 @@ namespace ARMeilleure.Instructions
             {
                 Operand ne = EmitVectorExtract(context, op.Rn, index, op.Size + 1, signedSrc);
 
-                Operand temp = EmitSatQ(context, ne, op.Size, signedSrc, signedDst);
+                Operand temp = signedSrc
+                    ? EmitSignedSrcSatQ(context, ne, op.Size, signedDst)
+                    : EmitUnsignedSrcSatQ(context, ne, op.Size, signedDst);
 
                 res = EmitVectorInsert(context, res, temp, part + index, op.Size);
             }
@@ -1528,74 +1533,248 @@ namespace ARMeilleure.Instructions
             context.Copy(d, res);
         }
 
-        // TSrc (16bit, 32bit, 64bit; signed, unsigned) > TDst (8bit, 16bit, 32bit; signed, unsigned).
-        public static Operand EmitSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedSrc, bool signedDst)
+        // TSrc (16bit, 32bit, 64bit; signed) > TDst (8bit, 16bit, 32bit; signed, unsigned).
+        // long SignedSrcSignedDstSatQ(long op, int size); ulong SignedSrcUnsignedDstSatQ(long op, int size);
+        public static Operand EmitSignedSrcSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedDst)
         {
-            if ((uint)sizeDst > 2u)
-            {
-                throw new ArgumentOutOfRangeException(nameof(sizeDst));
+            Debug.Assert(op.Type == OperandType.I64 && (uint)sizeDst <= 2u);
+
+            Operand lbl1 = Label();
+            Operand lblEnd = Label();
+
+            int eSize = 8 << sizeDst;
+
+            Operand maxT = signedDst ? Const((1L << (eSize - 1)) - 1L) : Const((1UL << eSize) - 1UL);
+            Operand minT = signedDst ? Const(-(1L << (eSize - 1))) : Const(0UL);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lbl1, res, maxT, Comparison.LessOrEqual);
+            context.Copy(res, maxT);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lblEnd, res, minT, Comparison.GreaterOrEqual);
+            context.Copy(res, minT);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-            MethodInfo info;
+        // TSrc (16bit, 32bit, 64bit; unsigned) > TDst (8bit, 16bit, 32bit; signed, unsigned).
+        // long UnsignedSrcSignedDstSatQ(ulong op, int size); ulong UnsignedSrcUnsignedDstSatQ(ulong op, int size);
+        public static Operand EmitUnsignedSrcSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedDst)
+        {
+            Debug.Assert(op.Type == OperandType.I64 && (uint)sizeDst <= 2u);
 
-            if (signedSrc)
-            {
-                info = signedDst
-                    ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedSrcSignedDstSatQ))
-                    : typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedSrcUnsignedDstSatQ));
-            }
-            else
-            {
-                info = signedDst
-                    ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedSrcSignedDstSatQ))
-                    : typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedSrcUnsignedDstSatQ));
+            Operand lblEnd = Label();
+
+            int eSize = 8 << sizeDst;
+
+            Operand maxL = signedDst ? Const((1L << (eSize - 1)) - 1L) : Const((1UL << eSize) - 1UL);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lblEnd, res, maxL, Comparison.LessOrEqualUI);
+            context.Copy(res, maxL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-            return context.Call(info, op, Const(sizeDst));
+        // long UnarySignedSatQAbsOrNeg(long op);
+        private static Operand EmitUnarySignedSatQAbsOrNeg(ArmEmitterContext context, Operand op)
+        {
+            Debug.Assert(op.Type == OperandType.I64);
+
+            Operand lblEnd = Label();
+
+            Operand minL = Const(long.MinValue);
+            Operand maxL = Const(long.MaxValue);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lblEnd, res, minL, Comparison.NotEqual);
+            context.Copy(res, maxL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-        // TSrc (64bit) == TDst (64bit); signed.
-        public static Operand EmitUnarySignedSatQAbsOrNeg(ArmEmitterContext context, Operand op)
+        // long BinarySignedSatQAdd(long op1, long op2);
+        public static Operand EmitBinarySignedSatQAdd(ArmEmitterContext context, Operand op1, Operand op2)
         {
-            Debug.Assert(((OpCodeSimd)context.CurrOp).Size == 3, "Invalid element size.");
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
 
-            return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnarySignedSatQAbsOrNeg)), op);
+            Operand lblEnd = Label();
+
+            Operand minL = Const(long.MinValue);
+            Operand maxL = Const(long.MaxValue);
+            Operand zero = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+
+            Operand left = context.BitwiseNot(context.BitwiseExclusiveOr(op1, op2));
+            Operand right = context.BitwiseExclusiveOr(op1, res);
+            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zero, Comparison.GreaterOrEqual);
+
+            Operand isPositive = context.ICompareGreaterOrEqual(op1, zero);
+            context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-        // TSrcs (64bit) == TDst (64bit); signed, unsigned.
-        public static Operand EmitBinarySatQAdd(ArmEmitterContext context, Operand op1, Operand op2, bool signed)
+        // ulong BinaryUnsignedSatQAdd(ulong op1, ulong op2);
+        public static Operand EmitBinaryUnsignedSatQAdd(ArmEmitterContext context, Operand op1, Operand op2)
         {
-            Debug.Assert(((OpCodeSimd)context.CurrOp).Size == 3, "Invalid element size.");
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
 
-            MethodInfo info = signed
-                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQAdd))
-                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQAdd));
+            Operand lblEnd = Label();
 
-            return context.Call(info, op1, op2);
+            Operand maxUL = Const(ulong.MaxValue);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+
+            context.BranchIf(lblEnd, res, op1, Comparison.GreaterOrEqualUI);
+            context.Copy(res, maxUL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-        // TSrcs (64bit) == TDst (64bit); signed, unsigned.
-        public static Operand EmitBinarySatQSub(ArmEmitterContext context, Operand op1, Operand op2, bool signed)
+        // long BinarySignedSatQSub(long op1, long op2);
+        public static Operand EmitBinarySignedSatQSub(ArmEmitterContext context, Operand op1, Operand op2)
         {
-            Debug.Assert(((OpCodeSimd)context.CurrOp).Size == 3, "Invalid element size.");
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
 
-            MethodInfo info = signed
-                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQSub))
-                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQSub));
+            Operand lblEnd = Label();
 
-            return context.Call(info, op1, op2);
+            Operand minL = Const(long.MinValue);
+            Operand maxL = Const(long.MaxValue);
+            Operand zero = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Subtract(op1, op2));
+
+            Operand left = context.BitwiseExclusiveOr(op1, op2);
+            Operand right = context.BitwiseExclusiveOr(op1, res);
+            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zero, Comparison.GreaterOrEqual);
+
+            Operand isPositive = context.ICompareGreaterOrEqual(op1, zero);
+            context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
-        // TSrcs (64bit) == TDst (64bit); signed, unsigned.
-        public static Operand EmitBinarySatQAccumulate(ArmEmitterContext context, Operand op1, Operand op2, bool signed)
+        // ulong BinaryUnsignedSatQSub(ulong op1, ulong op2);
+        public static Operand EmitBinaryUnsignedSatQSub(ArmEmitterContext context, Operand op1, Operand op2)
         {
-            Debug.Assert(((OpCodeSimd)context.CurrOp).Size == 3, "Invalid element size.");
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
 
-            MethodInfo info = signed
-                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQAcc))
-                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQAcc));
+            Operand lblEnd = Label();
 
-            return context.Call(info, op1, op2);
+            Operand zero = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Subtract(op1, op2));
+
+            context.BranchIf(lblEnd, op1, op2, Comparison.GreaterOrEqualUI);
+            context.Copy(res, zero);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // long BinarySignedSatQAcc(ulong op1, long op2);
+        private static Operand EmitBinarySignedSatQAcc(ArmEmitterContext context, Operand op1, Operand op2)
+        {
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
+
+            Operand lbl1 = Label();
+            Operand lbl2 = Label();
+            Operand lblEnd = Label();
+
+            Operand maxL = Const(long.MaxValue);
+            Operand zero = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+
+            context.BranchIf(lbl1, op1, maxL, Comparison.GreaterUI);
+            Operand notOp2AndRes = context.BitwiseAnd(context.BitwiseNot(op2), res);
+            context.BranchIf(lblEnd, notOp2AndRes, zero, Comparison.GreaterOrEqual);
+            context.Copy(res, maxL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lbl2, op2, zero, Comparison.Less);
+            context.Copy(res, maxL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl2);
+            context.BranchIf(lblEnd, res, maxL, Comparison.LessOrEqualUI);
+            context.Copy(res, maxL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // ulong BinaryUnsignedSatQAcc(long op1, ulong op2);
+        private static Operand EmitBinaryUnsignedSatQAcc(ArmEmitterContext context, Operand op1, Operand op2)
+        {
+            Debug.Assert(op1.Type == OperandType.I64 && op2.Type == OperandType.I64);
+
+            Operand lbl1 = Label();
+            Operand lblEnd = Label();
+
+            Operand maxUL = Const(ulong.MaxValue);
+            Operand maxL = Const(long.MaxValue);
+            Operand zero = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+
+            context.BranchIf(lbl1, op1, zero, Comparison.Less);
+            context.BranchIf(lblEnd, res, op1, Comparison.GreaterOrEqualUI);
+            context.Copy(res, maxUL);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lblEnd, op2, maxL, Comparison.GreaterUI);
+            context.BranchIf(lblEnd, res, zero, Comparison.GreaterOrEqual);
+            context.Copy(res, zero);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
         }
 
         public static Operand EmitFloatAbs(ArmEmitterContext context, Operand value, bool single, bool vector)

ARMeilleure/Instructions/InstEmitSimdHelper32.cs

@@ -219,6 +219,25 @@ namespace ARMeilleure.Instructions
 
         // Integer
 
+        public static void EmitVectorUnaryAccumulateOpI32(ArmEmitterContext context, Func1I emit, bool signed)
+        {
+            OpCode32Simd op = (OpCode32Simd)context.CurrOp;
+
+            Operand res = GetVecA32(op.Qd);
+
+            int elems = op.GetBytesCount() >> op.Size;
+
+            for (int index = 0; index < elems; index++)
+            {
+                Operand de = EmitVectorExtract32(context, op.Qd, op.Id + index, op.Size, signed);
+                Operand me = EmitVectorExtract32(context, op.Qm, op.Im + index, op.Size, signed);
+
+                res = EmitVectorInsert(context, res, context.Add(de, emit(me)), op.Id + index, op.Size);
+            }
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+
         public static void EmitVectorUnaryOpI32(ArmEmitterContext context, Func1I emit, bool signed)
         {
             OpCode32Simd op = (OpCode32Simd)context.CurrOp;
@@ -385,6 +404,18 @@ namespace ARMeilleure.Instructions
             EmitVectorUnaryOpI32(context, emit, true);
         }
 
+        public static void EmitVectorUnaryOpSx32(ArmEmitterContext context, Func1I emit, bool accumulate)
+        {
+            if (accumulate)
+            {
+                EmitVectorUnaryAccumulateOpI32(context, emit, true);
+            }
+            else
+            {
+                EmitVectorUnaryOpI32(context, emit, true);
+            }
+        }
+
         public static void EmitVectorBinaryOpSx32(ArmEmitterContext context, Func2I emit)
         {
             EmitVectorBinaryOpI32(context, emit, true);
@@ -400,6 +431,18 @@ namespace ARMeilleure.Instructions
             EmitVectorUnaryOpI32(context, emit, false);
         }
 
+        public static void EmitVectorUnaryOpZx32(ArmEmitterContext context, Func1I emit, bool accumulate)
+        {
+            if (accumulate)
+            {
+                EmitVectorUnaryAccumulateOpI32(context, emit, false);
+            }
+            else
+            {
+                EmitVectorUnaryOpI32(context, emit, false);
+            }
+        }
+
         public static void EmitVectorBinaryOpZx32(ArmEmitterContext context, Func2I emit)
         {
             EmitVectorBinaryOpI32(context, emit, false);
@@ -592,6 +635,34 @@ namespace ARMeilleure.Instructions
             context.Copy(GetVecA32(op.Qd), res);
         }
 
+        public static void EmitVectorPairwiseLongOpI32(ArmEmitterContext context, Func2I emit, bool signed)
+        {
+            OpCode32Simd op = (OpCode32Simd)context.CurrOp;
+
+            int elems = (op.Q ? 16 : 8) >> op.Size;
+            int pairs = elems >> 1;
+            int id = (op.Vd & 1) * pairs;
+
+            Operand res = GetVecA32(op.Qd);
+
+            for (int index = 0; index < pairs; index++)
+            {
+                int pairIndex = index << 1;
+                Operand m1 = EmitVectorExtract32(context, op.Qm, op.Im + pairIndex, op.Size, signed);
+                Operand m2 = EmitVectorExtract32(context, op.Qm, op.Im + pairIndex + 1, op.Size, signed);
+
+                if (op.Size == 2)
+                {
+                    m1 = signed ? context.SignExtend32(OperandType.I64, m1) : context.ZeroExtend32(OperandType.I64, m1);
+                    m2 = signed ? context.SignExtend32(OperandType.I64, m2) : context.ZeroExtend32(OperandType.I64, m2);
+                }
+
+                res = EmitVectorInsert(context, res, emit(m1, m2), id + index, op.Size + 1);
+            }
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+
         // Narrow
 
         public static void EmitVectorUnaryNarrowOp32(ArmEmitterContext context, Func1I emit, bool signed = false)

ARMeilleure/Instructions/InstEmitSimdShift.cs

@@ -1004,7 +1004,7 @@ namespace ARMeilleure.Instructions
                     e = EmitShrImm64(context, e, signedSrc, roundConst, shift); // shift <= 32
                 }
 
-                e = EmitSatQ(context, e, op.Size, signedSrc, signedDst);
+                e = signedSrc ? EmitSignedSrcSatQ(context, e, op.Size, signedDst) : EmitUnsignedSrcSatQ(context, e, op.Size, signedDst);
 
                 res = EmitVectorInsert(context, res, e, part + index, op.Size);
             }

ARMeilleure/Instructions/InstEmitSimdShift32.cs

@@ -33,64 +33,24 @@ namespace ARMeilleure.Instructions
             EmitShrImmSaturatingNarrowOp(context, op.U ? ShrImmSaturatingNarrowFlags.VectorZxZx : ShrImmSaturatingNarrowFlags.VectorSxSx);
         }
 
+        public static void Vqshrun(ArmEmitterContext context)
+        {
+            EmitShrImmSaturatingNarrowOp(context, ShrImmSaturatingNarrowFlags.VectorSxZx);
+        }
+
         public static void Vrshr(ArmEmitterContext context)
         {
-            OpCode32SimdShImm op = (OpCode32SimdShImm)context.CurrOp;
-            int shift = GetImmShr(op);
-            long roundConst = 1L << (shift - 1);
+            EmitRoundShrImmOp(context, accumulate: false);
+        }
 
-            if (op.U)
+        public static void Vrshrn(ArmEmitterContext context)
         {
-                if (op.Size < 2)
-                {
-                    EmitVectorUnaryOpZx32(context, (op1) =>
-                    {
-                        op1 = context.Add(op1, Const(op1.Type, roundConst));
+            EmitRoundShrImmNarrowOp(context, signed: false);
+        }
 
-                        return context.ShiftRightUI(op1, Const(shift));
-                    });
-                }
-                else if (op.Size == 2)
+        public static void Vrsra(ArmEmitterContext context)
         {
-                    EmitVectorUnaryOpZx32(context, (op1) =>
-                    {
-                        op1 = context.ZeroExtend32(OperandType.I64, op1);
-                        op1 = context.Add(op1, Const(op1.Type, roundConst));
-
-                        return context.ConvertI64ToI32(context.ShiftRightUI(op1, Const(shift)));
-                    });
-                }
-                else /* if (op.Size == 3) */
-                {
-                    EmitVectorUnaryOpZx32(context, (op1) => EmitShrImm64(context, op1, signed: false, roundConst, shift));
-                }
-            }
-            else
-            {
-                if (op.Size < 2)
-                {
-                    EmitVectorUnaryOpSx32(context, (op1) =>
-                    {
-                        op1 = context.Add(op1, Const(op1.Type, roundConst));
-
-                        return context.ShiftRightSI(op1, Const(shift));
-                    });
-                }
-                else if (op.Size == 2)
-                {
-                    EmitVectorUnaryOpSx32(context, (op1) =>
-                    {
-                        op1 = context.SignExtend32(OperandType.I64, op1);
-                        op1 = context.Add(op1, Const(op1.Type, roundConst));
-
-                        return context.ConvertI64ToI32(context.ShiftRightSI(op1, Const(shift)));
-                    });
-                }
-                else /* if (op.Size == 3) */
-                {
-                    EmitVectorUnaryOpZx32(context, (op1) => EmitShrImm64(context, op1, signed: true, roundConst, shift));
-                }
-            }
+            EmitRoundShrImmOp(context, accumulate: true);
         }
 
         public static void Vshl(ArmEmitterContext context)
@@ -191,6 +151,89 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static void EmitRoundShrImmOp(ArmEmitterContext context, bool accumulate)
+        {
+            OpCode32SimdShImm op = (OpCode32SimdShImm)context.CurrOp;
+            int shift = GetImmShr(op);
+            long roundConst = 1L << (shift - 1);
+
+            if (op.U)
+            {
+                if (op.Size < 2)
+                {
+                    EmitVectorUnaryOpZx32(context, (op1) =>
+                    {
+                        op1 = context.Add(op1, Const(op1.Type, roundConst));
+
+                        return context.ShiftRightUI(op1, Const(shift));
+                    }, accumulate);
+                }
+                else if (op.Size == 2)
+                {
+                    EmitVectorUnaryOpZx32(context, (op1) =>
+                    {
+                        op1 = context.ZeroExtend32(OperandType.I64, op1);
+                        op1 = context.Add(op1, Const(op1.Type, roundConst));
+
+                        return context.ConvertI64ToI32(context.ShiftRightUI(op1, Const(shift)));
+                    }, accumulate);
+                }
+                else /* if (op.Size == 3) */
+                {
+                    EmitVectorUnaryOpZx32(context, (op1) => EmitShrImm64(context, op1, signed: false, roundConst, shift), accumulate);
+                }
+            }
+            else
+            {
+                if (op.Size < 2)
+                {
+                    EmitVectorUnaryOpSx32(context, (op1) =>
+                    {
+                        op1 = context.Add(op1, Const(op1.Type, roundConst));
+
+                        return context.ShiftRightSI(op1, Const(shift));
+                    }, accumulate);
+                }
+                else if (op.Size == 2)
+                {
+                    EmitVectorUnaryOpSx32(context, (op1) =>
+                    {
+                        op1 = context.SignExtend32(OperandType.I64, op1);
+                        op1 = context.Add(op1, Const(op1.Type, roundConst));
+
+                        return context.ConvertI64ToI32(context.ShiftRightSI(op1, Const(shift)));
+                    }, accumulate);
+                }
+                else /* if (op.Size == 3) */
+                {
+                    EmitVectorUnaryOpZx32(context, (op1) => EmitShrImm64(context, op1, signed: true, roundConst, shift), accumulate);
+                }
+            }
+        }
+
+        private static void EmitRoundShrImmNarrowOp(ArmEmitterContext context, bool signed)
+        {
+            OpCode32SimdShImm op = (OpCode32SimdShImm)context.CurrOp;
+
+            int shift = GetImmShr(op);
+            long roundConst = 1L << (shift - 1);
+
+            EmitVectorUnaryNarrowOp32(context, (op1) =>
+            {
+                if (op.Size <= 1)
+                {
+                    op1 = context.Add(op1, Const(op1.Type, roundConst));
+                    op1 = signed ? context.ShiftRightSI(op1, Const(shift)) : context.ShiftRightUI(op1, Const(shift));
+                }
+                else /* if (op.Size == 2 && round) */
+                {
+                    op1 = EmitShrImm64(context, op1, signed, roundConst, shift); // shift <= 32
+                }
+
+                return op1;
+            }, signed);
+        }
+
         private static Operand EmitShlRegOp(ArmEmitterContext context, Operand op, Operand shiftLsB, int size, bool unsigned)
         {
             if (shiftLsB.Type == OperandType.I64)
@@ -289,7 +332,7 @@ namespace ARMeilleure.Instructions
                     op1 = EmitShrImm64(context, op1, signedSrc, roundConst, shift); // shift <= 32
                 }
 
-                return EmitSatQ(context, op1, 8 << op.Size, signedDst);
+                return EmitSatQ(context, op1, 8 << op.Size, signedSrc, signedDst);
             }, signedSrc);
         }
 
@@ -313,15 +356,20 @@ namespace ARMeilleure.Instructions
             return context.Call(info, value, Const(roundConst), Const(shift));
         }
 
-        private static Operand EmitSatQ(ArmEmitterContext context, Operand value, int eSize, bool signed)
+        private static Operand EmitSatQ(ArmEmitterContext context, Operand value, int eSize, bool signedSrc, bool signedDst)
         {
             Debug.Assert(eSize <= 32);
 
-            long intMin = signed ? -(1L << (eSize - 1)) : 0;
-            long intMax = signed ? (1L << (eSize - 1)) - 1 : (1L << eSize) - 1;
+            long intMin = signedDst ? -(1L << (eSize - 1)) : 0;
+            long intMax = signedDst ? (1L << (eSize - 1)) - 1 : (1L << eSize) - 1;
 
-            Operand gt = context.ICompareGreater(value, Const(value.Type, intMax));
-            Operand lt = context.ICompareLess(value, Const(value.Type, intMin));
+            Operand gt = signedSrc
+                ? context.ICompareGreater(value, Const(value.Type, intMax))
+                : context.ICompareGreaterUI(value, Const(value.Type, intMax));
+
+            Operand lt = signedSrc
+                ? context.ICompareLess(value, Const(value.Type, intMin))
+                : context.ICompareLessUI(value, Const(value.Type, intMin));
 
             value = context.ConditionalSelect(gt, Const(value.Type, intMax), value);
             value = context.ConditionalSelect(lt, Const(value.Type, intMin), value);

ARMeilleure/Instructions/InstEmitSystem32.cs

@@ -16,18 +16,13 @@ namespace ARMeilleure.Instructions
         {
             OpCode32System op = (OpCode32System)context.CurrOp;
 
-            if (op.Coproc != 15)
+            if (op.Coproc != 15 || op.Opc1 != 0)
             {
                 InstEmit.Und(context);
 
                 return;
             }
 
-            if (op.Opc1 != 0)
-            {
-                throw new NotImplementedException($"Unknown MRC Opc1 0x{op.Opc1:X16} at 0x{op.Address:X16}.");
-            }
-
             MethodInfo info;
 
             switch (op.CRn)
@@ -35,7 +30,7 @@ namespace ARMeilleure.Instructions
                 case 13: // Process and Thread Info.
                     if (op.CRm != 0)
                     {
-                        throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X16} at 0x{op.Address:X16}.");
+                        throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X} at 0x{op.Address:X} (0x{op.RawOpCode:X}).");
                     }
 
                     switch (op.Opc2)
@@ -44,7 +39,7 @@ namespace ARMeilleure.Instructions
                             info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl032)); break;
 
                         default:
-                            throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X16} at 0x{op.Address:X16}.");
+                            throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X} at 0x{op.Address:X} (0x{op.RawOpCode:X}).");
                     }
 
                     break;
@@ -59,11 +54,11 @@ namespace ARMeilleure.Instructions
                                     return; // No-op.
 
                                 default:
-                                    throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X16} at 0x{op.Address:X16}.");
+                                    throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X16} at 0x{op.Address:X16} (0x{op.RawOpCode:X}).");
                             }
 
                         default:
-                            throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X16} at 0x{op.Address:X16}.");
+                            throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X16} at 0x{op.Address:X16} (0x{op.RawOpCode:X}).");
                     }
 
                 default:
@@ -77,18 +72,13 @@ namespace ARMeilleure.Instructions
         {
             OpCode32System op = (OpCode32System)context.CurrOp;
 
-            if (op.Coproc != 15)
+            if (op.Coproc != 15 || op.Opc1 != 0)
             {
                 InstEmit.Und(context);
 
                 return;
             }
 
-            if (op.Opc1 != 0)
-            {
-                throw new NotImplementedException($"Unknown MRC Opc1 0x{op.Opc1:X16} at 0x{op.Address:X16}.");
-            }
-
             MethodInfo info;
 
             switch (op.CRn)
@@ -96,7 +86,7 @@ namespace ARMeilleure.Instructions
                 case 13: // Process and Thread Info.
                     if (op.CRm != 0)
                     {
-                        throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X16} at 0x{op.Address:X16}.");
+                        throw new NotImplementedException($"Unknown MRC CRm 0x{op.CRm:X} at 0x{op.Address:X} (0x{op.RawOpCode:X}).");
                     }
 
                     switch (op.Opc2)
@@ -108,13 +98,13 @@ namespace ARMeilleure.Instructions
                             info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidr32)); break;
 
                         default:
-                            throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X16} at 0x{op.Address:X16}.");
+                            throw new NotImplementedException($"Unknown MRC Opc2 0x{op.Opc2:X} at 0x{op.Address:X} (0x{op.RawOpCode:X}).");
                     }
 
                     break;
 
                 default:
-                    throw new NotImplementedException($"Unknown MRC 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
+                    throw new NotImplementedException($"Unknown MRC 0x{op.RawOpCode:X} at 0x{op.Address:X}.");
             }
 
             if (op.Rt == RegisterAlias.Aarch32Pc)
@@ -154,13 +144,13 @@ namespace ARMeilleure.Instructions
                             info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntpctEl0)); break;
 
                         default:
-                            throw new NotImplementedException($"Unknown MRRC Opc1 0x{opc:X16} at 0x{op.Address:X16}.");
+                            throw new NotImplementedException($"Unknown MRRC Opc1 0x{opc:X} at 0x{op.Address:X} (0x{op.RawOpCode:X}).");
                     }
 
                     break;
 
                 default:
-                    throw new NotImplementedException($"Unknown MRRC 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
+                    throw new NotImplementedException($"Unknown MRRC 0x{op.RawOpCode:X} at 0x{op.Address:X}.");
             }
 
             Operand result = context.Call(info);
@@ -265,7 +255,7 @@ namespace ARMeilleure.Instructions
                 case 0b1000: // FPEXC
                     throw new NotImplementedException("Supervisor Only");
                 default:
-                    throw new NotImplementedException($"Unknown VMRS 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
+                    throw new NotImplementedException($"Unknown VMRS 0x{op.RawOpCode:X} at 0x{op.Address:X}.");
             }
         }
 
@@ -288,7 +278,7 @@ namespace ARMeilleure.Instructions
                 case 0b1000: // FPEXC
                     throw new NotImplementedException("Supervisor Only");
                 default:
-                    throw new NotImplementedException($"Unknown VMSR 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
+                    throw new NotImplementedException($"Unknown VMSR 0x{op.RawOpCode:X} at 0x{op.Address:X}.");
             }
         }
 

ARMeilleure/Instructions/InstName.cs

@@ -545,6 +545,8 @@ namespace ARMeilleure.Instructions
         Strexh,
         Strh,
         Sxtb16,
+        Tbb,
+        Tbh,
         Teq,
         Trap,
         Tst,
@@ -601,6 +603,7 @@ namespace ARMeilleure.Instructions
         Vmin,
         Vminnm,
         Vmla,
+        Vmlal,
         Vmls,
         Vmlsl,
         Vmov,
@@ -618,15 +621,24 @@ namespace ARMeilleure.Instructions
         Vorn,
         Vorr,
         Vpadd,
+        Vpaddl,
         Vpmax,
         Vpmin,
+        Vqadd,
+        Vqdmulh,
+        Vqmovn,
+        Vqmovun,
         Vqrshrn,
         Vqrshrun,
         Vqshrn,
+        Vqshrun,
+        Vqsub,
         Vrev,
+        Vrhadd,
         Vrint,
         Vrintx,
         Vrshr,
+        Vrshrn,
         Vsel,
         Vshl,
         Vshll,
@@ -643,8 +655,10 @@ namespace ARMeilleure.Instructions
         Vrecps,
         Vrsqrte,
         Vrsqrts,
+        Vrsra,
         Vsra,
         Vsub,
+        Vsubl,
         Vsubw,
         Vtbl,
         Vtrn,

ARMeilleure/Instructions/SoftFallback.cs

@@ -91,7 +91,7 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (eSize != 64) */
                 {
-                    return SignedSrcSignedDstSatQ(value << shiftLsB, size);
+                    return SignedSrcSignedDstSatQ(value << shiftLsB, size); // InstEmitSimdHelper.EmitSignedSrcSatQ(signedDst: true).
                 }
             }
             else /* if (shiftLsB == 0) */
@@ -135,7 +135,7 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (eSize != 64) */
                 {
-                    return UnsignedSrcUnsignedDstSatQ(value << shiftLsB, size);
+                    return UnsignedSrcUnsignedDstSatQ(value << shiftLsB, size); // InstEmitSimdHelper.EmitUnsignedSrcSatQ(signedDst: false).
                 }
             }
             else /* if (shiftLsB == 0) */
@@ -509,7 +509,7 @@ namespace ARMeilleure.Instructions
 #endregion
 
 #region "Saturating"
-        public static long SignedSrcSignedDstSatQ(long op, int size)
+        private static long SignedSrcSignedDstSatQ(long op, int size)
         {
             ExecutionContext context = NativeInterface.GetContext();
 
@@ -536,54 +536,7 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        public static ulong SignedSrcUnsignedDstSatQ(long op, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            ulong tMaxValue = (1UL << eSize) - 1UL;
-            ulong tMinValue =  0UL;
-
-            if (op > (long)tMaxValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else if (op < (long)tMinValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMinValue;
-            }
-            else
-            {
-                return (ulong)op;
-            }
-        }
-
-        public static long UnsignedSrcSignedDstSatQ(ulong op, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            long tMaxValue = (1L << (eSize - 1)) - 1L;
-
-            if (op > (ulong)tMaxValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else
-            {
-                return (long)op;
-            }
-        }
-
-        public static ulong UnsignedSrcUnsignedDstSatQ(ulong op, int size)
+        private static ulong UnsignedSrcUnsignedDstSatQ(ulong op, int size)
         {
             ExecutionContext context = NativeInterface.GetContext();
 
@@ -602,208 +555,6 @@ namespace ARMeilleure.Instructions
                 return op;
             }
         }
-
-        public static long UnarySignedSatQAbsOrNeg(long op)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            if (op == long.MinValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return long.MaxValue;
-            }
-            else
-            {
-                return op;
-            }
-        }
-
-        public static long BinarySignedSatQAdd(long op1, long op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            long add = op1 + op2;
-
-            if ((~(op1 ^ op2) & (op1 ^ add)) < 0L)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                if (op1 < 0L)
-                {
-                    return long.MinValue;
-                }
-                else
-                {
-                    return long.MaxValue;
-                }
-            }
-            else
-            {
-                return add;
-            }
-        }
-
-        public static ulong BinaryUnsignedSatQAdd(ulong op1, ulong op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            ulong add = op1 + op2;
-
-            if ((add < op1) && (add < op2))
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return ulong.MaxValue;
-            }
-            else
-            {
-                return add;
-            }
-        }
-
-        public static long BinarySignedSatQSub(long op1, long op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            long sub = op1 - op2;
-
-            if (((op1 ^ op2) & (op1 ^ sub)) < 0L)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                if (op1 < 0L)
-                {
-                    return long.MinValue;
-                }
-                else
-                {
-                    return long.MaxValue;
-                }
-            }
-            else
-            {
-                return sub;
-            }
-        }
-
-        public static ulong BinaryUnsignedSatQSub(ulong op1, ulong op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            ulong sub = op1 - op2;
-
-            if (op1 < op2)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return ulong.MinValue;
-            }
-            else
-            {
-                return sub;
-            }
-        }
-
-        public static long BinarySignedSatQAcc(ulong op1, long op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            if (op1 <= (ulong)long.MaxValue)
-            {
-                // op1 from ulong.MinValue to (ulong)long.MaxValue
-                // op2 from long.MinValue to long.MaxValue
-
-                long add = (long)op1 + op2;
-
-                if ((~op2 & add) < 0L)
-                {
-                    context.Fpsr |= FPSR.Qc;
-
-                    return long.MaxValue;
-                }
-                else
-                {
-                    return add;
-                }
-            }
-            else if (op2 >= 0L)
-            {
-                // op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
-                // op2 from (long)ulong.MinValue to long.MaxValue
-
-                context.Fpsr |= FPSR.Qc;
-
-                return long.MaxValue;
-            }
-            else
-            {
-                // op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
-                // op2 from long.MinValue to (long)ulong.MinValue - 1L
-
-                ulong add = op1 + (ulong)op2;
-
-                if (add > (ulong)long.MaxValue)
-                {
-                    context.Fpsr |= FPSR.Qc;
-
-                    return long.MaxValue;
-                }
-                else
-                {
-                    return (long)add;
-                }
-            }
-        }
-
-        public static ulong BinaryUnsignedSatQAcc(long op1, ulong op2)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            if (op1 >= 0L)
-            {
-                // op1 from (long)ulong.MinValue to long.MaxValue
-                // op2 from ulong.MinValue to ulong.MaxValue
-
-                ulong add = (ulong)op1 + op2;
-
-                if ((add < (ulong)op1) && (add < op2))
-                {
-                    context.Fpsr |= FPSR.Qc;
-
-                    return ulong.MaxValue;
-                }
-                else
-                {
-                    return add;
-                }
-            }
-            else if (op2 > (ulong)long.MaxValue)
-            {
-                // op1 from long.MinValue to (long)ulong.MinValue - 1L
-                // op2 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
-
-                return (ulong)op1 + op2;
-            }
-            else
-            {
-                // op1 from long.MinValue to (long)ulong.MinValue - 1L
-                // op2 from ulong.MinValue to (ulong)long.MaxValue
-
-                long add = op1 + (long)op2;
-
-                if (add < (long)ulong.MinValue)
-                {
-                    context.Fpsr |= FPSR.Qc;
-
-                    return ulong.MinValue;
-                }
-                else
-                {
-                    return (ulong)add;
-                }
-            }
-        }
 #endregion
 
 #region "Count"

ARMeilleure/Translation/Delegates.cs

@@ -127,7 +127,7 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpcr)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpscr))); // A32 only.
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc))); // A32 only.
+            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl0)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl032))); // A32 only.
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SignalMemoryTracking)));
@@ -140,12 +140,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.WriteUInt64)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.WriteVector128)));
 
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQAcc)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQAdd)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinarySignedSatQSub)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQAcc)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQAdd)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.BinaryUnsignedSatQSub)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.CountLeadingSigns)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.CountLeadingZeros)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Crc32b)));
@@ -188,8 +182,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlReg)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlRegSatQ)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShrImm64)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedSrcSignedDstSatQ)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedSrcUnsignedDstSatQ)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbl1)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbl2)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbl3)));
@@ -198,12 +190,9 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx2)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx3)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx4)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnarySignedSatQAbsOrNeg)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlReg)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlRegSatQ)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShrImm64)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedSrcSignedDstSatQ)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedSrcUnsignedDstSatQ)));
 
             SetDelegateInfo(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)));
             SetDelegateInfo(typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert)));

ARMeilleure/Translation/PTC/Ptc.cs

@@ -27,7 +27,7 @@ namespace ARMeilleure.Translation.PTC
         private const string OuterHeaderMagicString = "PTCohd\0\0";
         private const string InnerHeaderMagicString = "PTCihd\0\0";
 
-        private const uint InternalVersion = 3585; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3683; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";

ARMeilleure/Translation/Translator.cs

@@ -44,15 +44,11 @@ namespace ARMeilleure.Translation
         private readonly IJitMemoryAllocator _allocator;
         private readonly ConcurrentQueue<KeyValuePair<ulong, TranslatedFunction>> _oldFuncs;
 
-        private readonly ConcurrentDictionary<ulong, object> _backgroundSet;
-        private readonly ConcurrentStack<RejitRequest> _backgroundStack;
-        private readonly AutoResetEvent _backgroundTranslatorEvent;
-        private readonly ReaderWriterLock _backgroundTranslatorLock;
-
         internal TranslatorCache<TranslatedFunction> Functions { get; }
         internal AddressTable<ulong> FunctionTable { get; }
         internal EntryTable<uint> CountTable { get; }
         internal TranslatorStubs Stubs { get; }
+        internal TranslatorQueue Queue { get; }
         internal IMemoryManager Memory { get; }
 
         private volatile int _threadCount;
@@ -67,10 +63,7 @@ namespace ARMeilleure.Translation
 
             _oldFuncs = new ConcurrentQueue<KeyValuePair<ulong, TranslatedFunction>>();
 
-            _backgroundSet = new ConcurrentDictionary<ulong, object>();
-            _backgroundStack = new ConcurrentStack<RejitRequest>();
-            _backgroundTranslatorEvent = new AutoResetEvent(false);
-            _backgroundTranslatorLock = new ReaderWriterLock();
+            Queue = new TranslatorQueue();
 
             JitCache.Initialize(allocator);
 
@@ -87,43 +80,6 @@ namespace ARMeilleure.Translation
             }
         }
 
-        private void TranslateStackedSubs()
-        {
-            while (_threadCount != 0)
-            {
-                _backgroundTranslatorLock.AcquireReaderLock(Timeout.Infinite);
-
-                if (_backgroundStack.TryPop(out RejitRequest request) &&
-                    _backgroundSet.TryRemove(request.Address, out _))
-                {
-                    TranslatedFunction func = Translate(request.Address, request.Mode, highCq: true);
-
-                    Functions.AddOrUpdate(request.Address, func.GuestSize, func, (key, oldFunc) =>
-                    {
-                        EnqueueForDeletion(key, oldFunc);
-                        return func;
-                    });
-
-                    if (PtcProfiler.Enabled)
-                    {
-                        PtcProfiler.UpdateEntry(request.Address, request.Mode, highCq: true);
-                    }
-
-                    RegisterFunction(request.Address, func);
-
-                    _backgroundTranslatorLock.ReleaseReaderLock();
-                }
-                else
-                {
-                    _backgroundTranslatorLock.ReleaseReaderLock();
-                    _backgroundTranslatorEvent.WaitOne();
-                }
-            }
-
-             // Wake up any other background translator threads, to encourage them to exit.
-            _backgroundTranslatorEvent.Set();
-        }
-
         public void Execute(State.ExecutionContext context, ulong address)
         {
             if (Interlocked.Increment(ref _threadCount) == 1)
@@ -155,7 +111,7 @@ namespace ARMeilleure.Translation
                 {
                     bool last = i != 0 && i == unboundedThreadCount - 1;
 
-                    Thread backgroundTranslatorThread = new Thread(TranslateStackedSubs)
+                    Thread backgroundTranslatorThread = new Thread(BackgroundTranslate)
                     {
                         Name = "CPU.BackgroundTranslatorThread." + i,
                         Priority = last ? ThreadPriority.Lowest : ThreadPriority.Normal
@@ -186,10 +142,9 @@ namespace ARMeilleure.Translation
 
             if (Interlocked.Decrement(ref _threadCount) == 0)
             {
-                _backgroundTranslatorEvent.Set();
-
                 ClearJitCache();
 
+                Queue.Dispose();
                 Stubs.Dispose();
                 FunctionTable.Dispose();
                 CountTable.Dispose();
@@ -317,6 +272,27 @@ namespace ARMeilleure.Translation
             return new TranslatedFunction(func, counter, funcSize, highCq);
         }
 
+        private void BackgroundTranslate()
+        {
+            while (_threadCount != 0 && Queue.TryDequeue(out RejitRequest request))
+            {
+                TranslatedFunction func = Translate(request.Address, request.Mode, highCq: true);
+
+                Functions.AddOrUpdate(request.Address, func.GuestSize, func, (key, oldFunc) =>
+                {
+                    EnqueueForDeletion(key, oldFunc);
+                    return func;
+                });
+
+                if (PtcProfiler.Enabled)
+                {
+                    PtcProfiler.UpdateEntry(request.Address, request.Mode, highCq: true);
+                }
+
+                RegisterFunction(request.Address, func);
+            }
+        }
+
         private struct Range
         {
             public ulong Start { get; }
@@ -504,11 +480,7 @@ namespace ARMeilleure.Translation
 
         internal void EnqueueForRejit(ulong guestAddress, ExecutionMode mode)
         {
-            if (_backgroundSet.TryAdd(guestAddress, null))
-            {
-                _backgroundStack.Push(new RejitRequest(guestAddress, mode));
-                _backgroundTranslatorEvent.Set();
-            }
+            Queue.Enqueue(guestAddress, mode);
         }
 
         private void EnqueueForDeletion(ulong guestAddress, TranslatedFunction func)
@@ -542,26 +514,23 @@ namespace ARMeilleure.Translation
 
         private void ClearRejitQueue(bool allowRequeue)
         {
-            _backgroundTranslatorLock.AcquireWriterLock(Timeout.Infinite);
-
-            if (allowRequeue)
+            if (!allowRequeue)
             {
-                while (_backgroundStack.TryPop(out var request))
+                Queue.Clear();
+
+                return;
+            }
+
+            lock (Queue.Sync)
+            {
+                while (Queue.Count > 0 && Queue.TryDequeue(out RejitRequest request))
                 {
                     if (Functions.TryGetValue(request.Address, out var func) && func.CallCounter != null)
                     {
                         Volatile.Write(ref func.CallCounter.Value, 0);
                     }
-
-                    _backgroundSet.TryRemove(request.Address, out _);
-                }
-            }
-            else
-            {
-                _backgroundStack.Clear();
-            }
-
-            _backgroundTranslatorLock.ReleaseWriterLock();
+                }
+            }
         }
     }
 }

ARMeilleure/Translation/TranslatorQueue.cs

@@ -0,0 +1,121 @@
+using ARMeilleure.Diagnostics;
+using ARMeilleure.State;
+using System;
+using System.Collections.Generic;
+using System.Threading;
+
+namespace ARMeilleure.Translation
+{
+    /// <summary>
+    /// Represents a queue of <see cref="RejitRequest"/>.
+    /// </summary>
+    /// <remarks>
+    /// This does not necessarily behave like a queue, i.e: a FIFO collection.
+    /// </remarks>
+    sealed class TranslatorQueue : IDisposable
+    {
+        private bool _disposed;
+        private readonly Stack<RejitRequest> _requests;
+        private readonly HashSet<ulong> _requestAddresses;
+
+        /// <summary>
+        /// Gets the object used to synchronize access to the <see cref="TranslatorQueue"/>.
+        /// </summary>
+        public object Sync { get; }
+
+        /// <summary>
+        /// Gets the number of requests in the <see cref="TranslatorQueue"/>.
+        /// </summary>
+        public int Count => _requests.Count;
+
+        /// <summary>
+        /// Initializes a new instance of the <see cref="TranslatorQueue"/> class.
+        /// </summary>
+        public TranslatorQueue()
+        {
+            Sync = new object();
+
+            _requests = new Stack<RejitRequest>();
+            _requestAddresses = new HashSet<ulong>();
+        }
+
+        /// <summary>
+        /// Enqueues a request with the specified <paramref name="address"/> and <paramref name="mode"/>.
+        /// </summary>
+        /// <param name="address">Address of request</param>
+        /// <param name="mode"><see cref="ExecutionMode"/> of request</param>
+        public void Enqueue(ulong address, ExecutionMode mode)
+        {
+            lock (Sync)
+            {
+                if (_requestAddresses.Add(address))
+                {
+                    _requests.Push(new RejitRequest(address, mode));
+
+                    TranslatorEventSource.Log.RejitQueueAdd(1);
+
+                    Monitor.Pulse(Sync);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Tries to dequeue a <see cref="RejitRequest"/>. This will block the thread until a <see cref="RejitRequest"/>
+        /// is enqueued or the <see cref="TranslatorQueue"/> is disposed.
+        /// </summary>
+        /// <param name="result"><see cref="RejitRequest"/> dequeued</param>
+        /// <returns><see langword="true"/> on success; otherwise <see langword="false"/></returns>
+        public bool TryDequeue(out RejitRequest result)
+        {
+            while (!_disposed)
+            {
+                lock (Sync)
+                {
+                    if (_requests.TryPop(out result))
+                    {
+                        _requestAddresses.Remove(result.Address);
+
+                        TranslatorEventSource.Log.RejitQueueAdd(-1);
+
+                        return true;
+                    }
+
+                    Monitor.Wait(Sync);
+                }
+            }
+
+            result = default;
+
+            return false;
+        }
+
+        /// <summary>
+        /// Clears the <see cref="TranslatorQueue"/>.
+        /// </summary>
+        public void Clear()
+        {
+            lock (Sync)
+            {
+                TranslatorEventSource.Log.RejitQueueAdd(-_requests.Count);
+
+                _requests.Clear();
+                _requestAddresses.Clear();
+
+                Monitor.PulseAll(Sync);
+            }
+        }
+
+        /// <summary>
+        /// Releases all resources used by the <see cref="TranslatorQueue"/> instance.
+        /// </summary>
+        public void Dispose()
+        {
+            if (!_disposed)
+            {
+                _disposed = true;
+
+                Clear();
+            }
+        }
+    }
+}

Ryujinx.Audio/Renderer/Server/AudioRenderSystem.cs

@@ -55,7 +55,6 @@ namespace Ryujinx.Audio.Renderer.Server
         private uint _processHandle;
         private ulong _appletResourceId;
 
-        private WritableRegion _workBufferRegion;
         private MemoryHandle _workBufferMemoryPin;
 
         private Memory<float> _mixBuffer;
@@ -98,7 +97,15 @@ namespace Ryujinx.Audio.Renderer.Server
             _sessionId = 0;
         }
 
-        public ResultCode Initialize(ref AudioRendererConfiguration parameter, uint processHandle, CpuAddress workBuffer, ulong workBufferSize, int sessionId, ulong appletResourceId, IVirtualMemoryManager memoryManager)
+        public ResultCode Initialize(
+            ref AudioRendererConfiguration parameter,
+            uint processHandle,
+            Memory<byte> workBufferMemory,
+            CpuAddress workBuffer,
+            ulong workBufferSize,
+            int sessionId,
+            ulong appletResourceId,
+            IVirtualMemoryManager memoryManager)
         {
             if (!BehaviourContext.CheckValidRevision(parameter.Revision))
             {
@@ -134,11 +141,10 @@ namespace Ryujinx.Audio.Renderer.Server
 
             WorkBufferAllocator workBufferAllocator;
 
-            _workBufferRegion = MemoryManager.GetWritableRegion(workBuffer, (int)workBufferSize);
-            _workBufferRegion.Memory.Span.Fill(0);
-            _workBufferMemoryPin = _workBufferRegion.Memory.Pin();
+            workBufferMemory.Span.Fill(0);
+            _workBufferMemoryPin = workBufferMemory.Pin();
 
-            workBufferAllocator = new WorkBufferAllocator(_workBufferRegion.Memory);
+            workBufferAllocator = new WorkBufferAllocator(workBufferMemory);
 
             PoolMapper poolMapper = new PoolMapper(processHandle, false);
             poolMapper.InitializeSystemPool(ref _dspMemoryPoolState, workBuffer, workBufferSize);
@@ -841,7 +847,6 @@ namespace Ryujinx.Audio.Renderer.Server
                 _manager.Unregister(this);
                 _terminationEvent.Dispose();
                 _workBufferMemoryPin.Dispose();
-                _workBufferRegion.Dispose();
 
                 if (MemoryManager is IRefCounted rc)
                 {

Ryujinx.Audio/Renderer/Server/AudioRendererManager.cs

@@ -305,13 +305,34 @@ namespace Ryujinx.Audio.Renderer.Server
         /// <param name="workBufferSize">The guest work buffer size.</param>
         /// <param name="processHandle">The process handle of the application.</param>
         /// <returns>A <see cref="ResultCode"/> reporting an error or a success.</returns>
-        public ResultCode OpenAudioRenderer(out AudioRenderSystem renderer, IVirtualMemoryManager memoryManager, ref AudioRendererConfiguration parameter, ulong appletResourceUserId, ulong workBufferAddress, ulong workBufferSize, uint processHandle, float volume)
+        public ResultCode OpenAudioRenderer(
+            out AudioRenderSystem renderer,
+            IVirtualMemoryManager memoryManager,
+            ref AudioRendererConfiguration parameter,
+            ulong appletResourceUserId,
+            ulong workBufferAddress,
+            ulong workBufferSize,
+            uint processHandle,
+            float volume)
         {
             int sessionId = AcquireSessionId();
 
             AudioRenderSystem audioRenderer = new AudioRenderSystem(this, _sessionsSystemEvent[sessionId]);
 
-            ResultCode result = audioRenderer.Initialize(ref parameter, processHandle, workBufferAddress, workBufferSize, sessionId, appletResourceUserId, memoryManager);
+            // TODO: Eventually, we should try to use the guest supplied work buffer instead of allocating
+            // our own. However, it was causing problems on some applications that would unmap the memory
+            // before the audio renderer was fully disposed.
+            Memory<byte> workBufferMemory = GC.AllocateArray<byte>((int)workBufferSize, pinned: true);
+
+            ResultCode result = audioRenderer.Initialize(
+                ref parameter,
+                processHandle,
+                workBufferMemory,
+                workBufferAddress,
+                workBufferSize,
+                sessionId,
+                appletResourceUserId,
+                memoryManager);
 
             if (result == ResultCode.Success)
             {

Ryujinx.Graphics.GAL/Format.cs

@@ -151,6 +151,190 @@ namespace Ryujinx.Graphics.GAL
 
     public static class FormatExtensions
     {
+        /// <summary>
+        /// The largest scalar size for a buffer format.
+        /// </summary>
+        public const int MaxBufferFormatScalarSize = 4;
+
+        /// <summary>
+        /// Gets the byte size for a single component of this format, or its packed size.
+        /// </summary>
+        /// <param name="format">Texture format</param>
+        /// <returns>Byte size for a single component, or packed size</returns>
+        public static int GetScalarSize(this Format format)
+        {
+            switch (format)
+            {
+                case Format.R8Unorm:
+                case Format.R8Snorm:
+                case Format.R8Uint:
+                case Format.R8Sint:
+                case Format.R8G8Unorm:
+                case Format.R8G8Snorm:
+                case Format.R8G8Uint:
+                case Format.R8G8Sint:
+                case Format.R8G8B8Unorm:
+                case Format.R8G8B8Snorm:
+                case Format.R8G8B8Uint:
+                case Format.R8G8B8Sint:
+                case Format.R8G8B8A8Unorm:
+                case Format.R8G8B8A8Snorm:
+                case Format.R8G8B8A8Uint:
+                case Format.R8G8B8A8Sint:
+                case Format.R8G8B8A8Srgb:
+                case Format.R4G4Unorm:
+                case Format.R8Uscaled:
+                case Format.R8Sscaled:
+                case Format.R8G8Uscaled:
+                case Format.R8G8Sscaled:
+                case Format.R8G8B8Uscaled:
+                case Format.R8G8B8Sscaled:
+                case Format.R8G8B8A8Uscaled:
+                case Format.R8G8B8A8Sscaled:
+                case Format.B8G8R8A8Unorm:
+                case Format.B8G8R8A8Srgb:
+                    return 1;
+
+                case Format.R16Float:
+                case Format.R16Unorm:
+                case Format.R16Snorm:
+                case Format.R16Uint:
+                case Format.R16Sint:
+                case Format.R16G16Float:
+                case Format.R16G16Unorm:
+                case Format.R16G16Snorm:
+                case Format.R16G16Uint:
+                case Format.R16G16Sint:
+                case Format.R16G16B16Float:
+                case Format.R16G16B16Unorm:
+                case Format.R16G16B16Snorm:
+                case Format.R16G16B16Uint:
+                case Format.R16G16B16Sint:
+                case Format.R16G16B16A16Float:
+                case Format.R16G16B16A16Unorm:
+                case Format.R16G16B16A16Snorm:
+                case Format.R16G16B16A16Uint:
+                case Format.R16G16B16A16Sint:
+                case Format.R4G4B4A4Unorm:
+                case Format.R5G5B5X1Unorm:
+                case Format.R5G5B5A1Unorm:
+                case Format.R5G6B5Unorm:
+                case Format.R16Uscaled:
+                case Format.R16Sscaled:
+                case Format.R16G16Uscaled:
+                case Format.R16G16Sscaled:
+                case Format.R16G16B16Uscaled:
+                case Format.R16G16B16Sscaled:
+                case Format.R16G16B16A16Uscaled:
+                case Format.R16G16B16A16Sscaled:
+                case Format.B5G6R5Unorm:
+                case Format.B5G5R5A1Unorm:
+                case Format.A1B5G5R5Unorm:
+                    return 2;
+
+                case Format.R32Float:
+                case Format.R32Uint:
+                case Format.R32Sint:
+                case Format.R32G32Float:
+                case Format.R32G32Uint:
+                case Format.R32G32Sint:
+                case Format.R32G32B32Float:
+                case Format.R32G32B32Uint:
+                case Format.R32G32B32Sint:
+                case Format.R32G32B32A32Float:
+                case Format.R32G32B32A32Uint:
+                case Format.R32G32B32A32Sint:
+                case Format.R10G10B10A2Unorm:
+                case Format.R10G10B10A2Uint:
+                case Format.R11G11B10Float:
+                case Format.R9G9B9E5Float:
+                case Format.R32Uscaled:
+                case Format.R32Sscaled:
+                case Format.R32G32Uscaled:
+                case Format.R32G32Sscaled:
+                case Format.R32G32B32Uscaled:
+                case Format.R32G32B32Sscaled:
+                case Format.R32G32B32A32Uscaled:
+                case Format.R32G32B32A32Sscaled:
+                case Format.R10G10B10A2Snorm:
+                case Format.R10G10B10A2Sint:
+                case Format.R10G10B10A2Uscaled:
+                case Format.R10G10B10A2Sscaled:
+                    return 4;
+
+                case Format.S8Uint:
+                    return 1;
+                case Format.D16Unorm:
+                    return 2;
+                case Format.S8UintD24Unorm:
+                case Format.D32Float:
+                case Format.D24UnormS8Uint:
+                    return 4;
+                case Format.D32FloatS8Uint:
+                    return 8;
+
+                case Format.Bc1RgbaUnorm:
+                case Format.Bc1RgbaSrgb:
+                    return 8;
+
+                case Format.Bc2Unorm:
+                case Format.Bc3Unorm:
+                case Format.Bc2Srgb:
+                case Format.Bc3Srgb:
+                case Format.Bc4Unorm:
+                case Format.Bc4Snorm:
+                case Format.Bc5Unorm:
+                case Format.Bc5Snorm:
+                case Format.Bc7Unorm:
+                case Format.Bc7Srgb:
+                case Format.Bc6HSfloat:
+                case Format.Bc6HUfloat:
+                    return 16;
+
+                case Format.Etc2RgbUnorm:
+                case Format.Etc2RgbPtaUnorm:
+                case Format.Etc2RgbSrgb:
+                case Format.Etc2RgbPtaSrgb:
+                    return 8;
+
+                case Format.Etc2RgbaUnorm:
+                case Format.Etc2RgbaSrgb:
+                    return 16;
+
+                case Format.Astc4x4Unorm:
+                case Format.Astc5x4Unorm:
+                case Format.Astc5x5Unorm:
+                case Format.Astc6x5Unorm:
+                case Format.Astc6x6Unorm:
+                case Format.Astc8x5Unorm:
+                case Format.Astc8x6Unorm:
+                case Format.Astc8x8Unorm:
+                case Format.Astc10x5Unorm:
+                case Format.Astc10x6Unorm:
+                case Format.Astc10x8Unorm:
+                case Format.Astc10x10Unorm:
+                case Format.Astc12x10Unorm:
+                case Format.Astc12x12Unorm:
+                case Format.Astc4x4Srgb:
+                case Format.Astc5x4Srgb:
+                case Format.Astc5x5Srgb:
+                case Format.Astc6x5Srgb:
+                case Format.Astc6x6Srgb:
+                case Format.Astc8x5Srgb:
+                case Format.Astc8x6Srgb:
+                case Format.Astc8x8Srgb:
+                case Format.Astc10x5Srgb:
+                case Format.Astc10x6Srgb:
+                case Format.Astc10x8Srgb:
+                case Format.Astc10x10Srgb:
+                case Format.Astc12x10Srgb:
+                case Format.Astc12x12Srgb:
+                    return 16;
+            }
+
+            return 1;
+        }
+
         /// <summary>
         /// Checks if the texture format is valid to use as image format.
         /// </summary>

Ryujinx.Graphics.Gpu/Image/TextureBindingsManager.cs

@@ -47,6 +47,7 @@ namespace Ryujinx.Graphics.Gpu.Image
 
             public int TextureHandle;
             public int SamplerHandle;
+            public Format ImageFormat;
             public int InvalidatedSequence;
             public Texture CachedTexture;
             public Sampler CachedSampler;
@@ -564,6 +565,9 @@ namespace Ryujinx.Graphics.Gpu.Image
                     // Buffers are frequently re-created to accomodate larger data, so we need to re-bind
                     // to ensure we're not using a old buffer that was already deleted.
                     _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, texture.Range.GetSubRange(0).Address, texture.Size, bindingInfo, bindingInfo.Format, false);
+
+                    // Cache is not used for buffer texture, it must always rebind.
+                    state.CachedTexture = null;
                 }
                 else
                 {
@@ -659,14 +663,16 @@ namespace Ryujinx.Graphics.Gpu.Image
                         cachedTexture?.SignalModified();
                     }
 
-                    if ((usageFlags & TextureUsageFlags.NeedsScaleValue) != 0 &&
-                        UpdateScale(state.CachedTexture, usageFlags, scaleIndex, stage))
+                    Format format = bindingInfo.Format == 0 ? cachedTexture.Format : bindingInfo.Format;
+
+                    if (state.ImageFormat != format ||
+                        ((usageFlags & TextureUsageFlags.NeedsScaleValue) != 0 &&
+                        UpdateScale(state.CachedTexture, usageFlags, scaleIndex, stage)))
                     {
                         ITexture hostTextureRebind = state.CachedTexture.GetTargetTexture(bindingInfo.Target);
 
-                        Format format = bindingInfo.Format == 0 ? cachedTexture.Format : bindingInfo.Format;
-
                         state.Texture = hostTextureRebind;
+                        state.ImageFormat = format;
 
                         _context.Renderer.Pipeline.SetImage(bindingInfo.Binding, hostTextureRebind, format);
                     }
@@ -696,6 +702,9 @@ namespace Ryujinx.Graphics.Gpu.Image
                     }
 
                     _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, texture.Range.GetSubRange(0).Address, texture.Size, bindingInfo, format, true);
+
+                    // Cache is not used for buffer texture, it must always rebind.
+                    state.CachedTexture = null;
                 }
                 else
                 {
@@ -721,6 +730,8 @@ namespace Ryujinx.Graphics.Gpu.Image
                             format = texture.Format;
                         }
 
+                        state.ImageFormat = format;
+
                         _context.Renderer.Pipeline.SetImage(bindingInfo.Binding, hostTexture, format);
                     }
 

Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs

@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
         private const ushort FileFormatVersionMajor = 1;
         private const ushort FileFormatVersionMinor = 2;
         private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 3525;
+        private const uint CodeGenVersion = 3672;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

Ryujinx.Graphics.Shader/Translation/Optimizations/ConstantFolding.cs

@@ -153,6 +153,10 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                     EvaluateFPUnary(operation, (x) => float.IsNaN(x));
                     break;
 
+                case Instruction.LoadConstant:
+                    operation.TurnIntoCopy(Cbuf(operation.GetSource(0).Value, operation.GetSource(1).Value));
+                    break;
+
                 case Instruction.Maximum:
                     EvaluateBinary(operation, (x, y) => Math.Max(x, y));
                     break;

Ryujinx.Graphics.Vulkan/BufferHolder.cs

@@ -27,7 +27,7 @@ namespace Ryujinx.Graphics.Vulkan
         private readonly Auto<MemoryAllocation> _allocationAuto;
         private readonly ulong _bufferHandle;
 
-        private CacheByRange<BufferHolder> _cachedConvertedIndexBuffers;
+        private CacheByRange<BufferHolder> _cachedConvertedBuffers;
 
         public int Size { get; }
 
@@ -109,7 +109,7 @@ namespace Ryujinx.Graphics.Vulkan
         {
             if (isWrite)
             {
-                _cachedConvertedIndexBuffers.Clear();
+                _cachedConvertedBuffers.Clear();
             }
 
             return _buffer;
@@ -364,13 +364,35 @@ namespace Ryujinx.Graphics.Vulkan
 
         public Auto<DisposableBuffer> GetBufferI8ToI16(CommandBufferScoped cbs, int offset, int size)
         {
-            if (!_cachedConvertedIndexBuffers.TryGetValue(offset, size, out var holder))
+            var key = new I8ToI16CacheKey();
+
+            if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
             {
                 holder = _gd.BufferManager.Create(_gd, (size * 2 + 3) & ~3);
 
                 _gd.HelperShader.ConvertI8ToI16(_gd, cbs, this, holder, offset, size);
 
-                _cachedConvertedIndexBuffers.Add(offset, size, holder);
+                _cachedConvertedBuffers.Add(offset, size, key, holder);
+            }
+
+            return holder.GetBuffer();
+        }
+
+        public Auto<DisposableBuffer> GetAlignedVertexBuffer(CommandBufferScoped cbs, int offset, int size, int stride, int alignment)
+        {
+            var key = new AlignedVertexBufferCacheKey(_gd, stride, alignment);
+
+            if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
+            {
+                int alignedStride = (stride + (alignment - 1)) & -alignment;
+
+                holder = _gd.BufferManager.Create(_gd, (size / stride) * alignedStride);
+
+                _gd.HelperShader.ChangeStride(_gd, cbs, this, holder, offset, size, stride, alignedStride);
+
+                key.SetBuffer(holder.GetBuffer());
+
+                _cachedConvertedBuffers.Add(offset, size, key, holder);
             }
 
             return holder.GetBuffer();
@@ -382,7 +404,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             _buffer.Dispose();
             _allocationAuto.Dispose();
-            _cachedConvertedIndexBuffers.Dispose();
+            _cachedConvertedBuffers.Dispose();
         }
     }
 }

Ryujinx.Graphics.Vulkan/BufferManager.cs

@@ -130,6 +130,16 @@ namespace Ryujinx.Graphics.Vulkan
             return null;
         }
 
+        public Auto<DisposableBuffer> GetAlignedVertexBuffer(CommandBufferScoped cbs, BufferHandle handle, int offset, int size, int stride, int alignment)
+        {
+            if (TryGetBuffer(handle, out var holder))
+            {
+                return holder.GetAlignedVertexBuffer(cbs, offset, size, stride, alignment);
+            }
+
+            return null;
+        }
+
         public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, BufferHandle handle, bool isWrite, out int size)
         {
             if (TryGetBuffer(handle, out var holder))

Ryujinx.Graphics.Vulkan/BufferState.cs

@@ -7,28 +7,28 @@ namespace Ryujinx.Graphics.Vulkan
     {
         public static BufferState Null => new BufferState(null, 0, 0);
 
-        private readonly Auto<DisposableBuffer> _buffer;
         private readonly int _offset;
         private readonly int _size;
-        private readonly ulong _stride;
         private readonly IndexType _type;
 
+        private readonly Auto<DisposableBuffer> _buffer;
+
         public BufferState(Auto<DisposableBuffer> buffer, int offset, int size, IndexType type)
         {
             _buffer = buffer;
+
             _offset = offset;
             _size = size;
-            _stride = 0;
             _type = type;
             buffer?.IncrementReferenceCount();
         }
 
-        public BufferState(Auto<DisposableBuffer> buffer, int offset, int size, ulong stride = 0UL)
+        public BufferState(Auto<DisposableBuffer> buffer, int offset, int size)
         {
             _buffer = buffer;
+
             _offset = offset;
             _size = size;
-            _stride = stride;
             _type = IndexType.Uint16;
             buffer?.IncrementReferenceCount();
         }
@@ -51,30 +51,6 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
-        public void BindVertexBuffer(VulkanRenderer gd, CommandBufferScoped cbs, uint binding)
-        {
-            if (_buffer != null)
-            {
-                var buffer = _buffer.Get(cbs, _offset, _size).Value;
-
-                if (gd.Capabilities.SupportsExtendedDynamicState)
-                {
-                    gd.ExtendedDynamicStateApi.CmdBindVertexBuffers2(
-                        cbs.CommandBuffer,
-                        binding,
-                        1,
-                        buffer,
-                        (ulong)_offset,
-                        (ulong)_size,
-                        _stride);
-                }
-                else
-                {
-                    gd.Api.CmdBindVertexBuffers(cbs.CommandBuffer, binding, 1, buffer, (ulong)_offset);
-                }
-            }
-        }
-
         public void Dispose()
         {
             _buffer?.DecrementReferenceCount();

Ryujinx.Graphics.Vulkan/CacheByRange.cs

@@ -3,29 +3,110 @@ using System.Collections.Generic;
 
 namespace Ryujinx.Graphics.Vulkan
 {
-    struct CacheByRange<T> where T : IDisposable
+    interface ICacheKey : IDisposable
     {
-        private Dictionary<ulong, T> _ranges;
-
-        public void Add(int offset, int size, T value)
-        {
-            EnsureInitialized();
-            _ranges.Add(PackRange(offset, size),  value);
+        bool KeyEqual(ICacheKey other);
     }
 
-        public bool TryGetValue(int offset, int size, out T value)
+    struct I8ToI16CacheKey : ICacheKey
     {
-            EnsureInitialized();
-            return _ranges.TryGetValue(PackRange(offset, size), out value);
+        public I8ToI16CacheKey() { }
+
+        public bool KeyEqual(ICacheKey other)
+        {
+            return other is I8ToI16CacheKey;
+        }
+
+        public void Dispose() { }
+    }
+
+    struct AlignedVertexBufferCacheKey : ICacheKey
+    {
+        private readonly int _stride;
+        private readonly int _alignment;
+
+        // Used to notify the pipeline that bindings have invalidated on dispose.
+        private readonly VulkanRenderer _gd;
+        private Auto<DisposableBuffer> _buffer;
+
+        public AlignedVertexBufferCacheKey(VulkanRenderer gd, int stride, int alignment)
+        {
+            _gd = gd;
+            _stride = stride;
+            _alignment = alignment;
+            _buffer = null;
+        }
+
+        public bool KeyEqual(ICacheKey other)
+        {
+            return other is AlignedVertexBufferCacheKey entry &&
+                entry._stride == _stride &&
+                entry._alignment == _alignment;
+        }
+
+        public void SetBuffer(Auto<DisposableBuffer> buffer)
+        {
+            _buffer = buffer;
+        }
+
+        public void Dispose()
+        {
+            _gd.PipelineInternal.DirtyVertexBuffer(_buffer);
+        }
+    }
+
+    struct CacheByRange<T> where T : IDisposable
+    {
+        private struct Entry
+        {
+            public ICacheKey Key;
+            public T Value;
+
+            public Entry(ICacheKey key, T value)
+            {
+                Key = key;
+                Value = value;
+            }
+        }
+
+        private Dictionary<ulong, List<Entry>> _ranges;
+
+        public void Add(int offset, int size, ICacheKey key, T value)
+        {
+            List<Entry> entries = GetEntries(offset, size);
+
+            entries.Add(new Entry(key, value));
+        }
+
+        public bool TryGetValue(int offset, int size, ICacheKey key, out T value)
+        {
+            List<Entry> entries = GetEntries(offset, size);
+
+            foreach (Entry entry in entries)
+            {
+                if (entry.Key.KeyEqual(key))
+                {
+                    value = entry.Value;
+
+                    return true;
+                }
+            }
+
+            value = default;
+            return false;
         }
 
         public void Clear()
         {
             if (_ranges != null)
             {
-                foreach (T value in _ranges.Values)
+                foreach (List<Entry> entries in _ranges.Values)
                 {
-                    value.Dispose();
+                    foreach (Entry entry in entries)
+                    {
+                        entry.Key.Dispose();
+                        entry.Value.Dispose();
+                    }
                 }
 
                 _ranges.Clear();
@@ -33,12 +114,23 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
-        private void EnsureInitialized()
+        private List<Entry> GetEntries(int offset, int size)
         {
             if (_ranges == null)
             {
-                _ranges = new Dictionary<ulong, T>();
+                _ranges = new Dictionary<ulong, List<Entry>>();
             }
+
+            ulong key = PackRange(offset, size);
+
+            List<Entry> value;
+            if (!_ranges.TryGetValue(key, out value))
+            {
+                value = new List<Entry>();
+                _ranges.Add(key, value);
+            }
+
+            return value;
         }
 
         private static ulong PackRange(int offset, int size)

Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs

@@ -185,6 +185,34 @@ namespace Ryujinx.Graphics.Vulkan
             SignalDirty(DirtyFlags.Storage);
         }
 
+        public void SetStorageBuffers(CommandBuffer commandBuffer, int first, ReadOnlySpan<Auto<DisposableBuffer>> buffers)
+        {
+            for (int i = 0; i < buffers.Length; i++)
+            {
+                var vkBuffer = buffers[i];
+                int index = first + i;
+
+                ref Auto<DisposableBuffer> currentVkBuffer = ref _storageBufferRefs[index];
+
+                DescriptorBufferInfo info = new DescriptorBufferInfo()
+                {
+                    Offset = 0,
+                    Range = Vk.WholeSize
+                };
+                ref DescriptorBufferInfo currentInfo = ref _storageBuffers[index];
+
+                if (vkBuffer != currentVkBuffer || currentInfo.Offset != info.Offset || currentInfo.Range != info.Range)
+                {
+                    _storageSet[index] = false;
+
+                    currentInfo = info;
+                    currentVkBuffer = vkBuffer;
+                }
+            }
+
+            SignalDirty(DirtyFlags.Storage);
+        }
+
         public void SetTextureAndSampler(CommandBufferScoped cbs, ShaderStage stage, int binding, ITexture texture, ISampler sampler)
         {
             if (texture == null)
@@ -388,8 +416,15 @@ namespace Ryujinx.Graphics.Vulkan
                         }
 
                         ReadOnlySpan<DescriptorBufferInfo> storageBuffers = _storageBuffers;
+                        if (program.HasMinimalLayout)
+                        {
+                            dsc.UpdateBuffers(0, binding, storageBuffers.Slice(binding, count), DescriptorType.StorageBuffer);
+                        }
+                        else
+                        {
                             dsc.UpdateStorageBuffers(0, binding, storageBuffers.Slice(binding, count));
                         }
+                    }
                     else if (setIndex == PipelineBase.TextureSetIndex)
                     {
                         if (((uint)binding % (Constants.MaxTexturesPerStage * 2)) < Constants.MaxTexturesPerStage || program.HasMinimalLayout)

Ryujinx.Graphics.Vulkan/HardwareCapabilities.cs

@@ -10,6 +10,7 @@ namespace Ryujinx.Graphics.Vulkan
         public readonly bool SupportsFragmentShaderInterlock;
         public readonly bool SupportsGeometryShaderPassthrough;
         public readonly bool SupportsSubgroupSizeControl;
+        public readonly bool SupportsShaderInt8;
         public readonly bool SupportsConditionalRendering;
         public readonly bool SupportsExtendedDynamicState;
         public readonly bool SupportsMultiView;
@@ -29,6 +30,7 @@ namespace Ryujinx.Graphics.Vulkan
             bool supportsFragmentShaderInterlock,
             bool supportsGeometryShaderPassthrough,
             bool supportsSubgroupSizeControl,
+            bool supportsShaderInt8,
             bool supportsConditionalRendering,
             bool supportsExtendedDynamicState,
             bool supportsMultiView,
@@ -47,6 +49,7 @@ namespace Ryujinx.Graphics.Vulkan
             SupportsFragmentShaderInterlock = supportsFragmentShaderInterlock;
             SupportsGeometryShaderPassthrough = supportsGeometryShaderPassthrough;
             SupportsSubgroupSizeControl = supportsSubgroupSizeControl;
+            SupportsShaderInt8 = supportsShaderInt8;
             SupportsConditionalRendering = supportsConditionalRendering;
             SupportsExtendedDynamicState = supportsExtendedDynamicState;
             SupportsMultiView = supportsMultiView;

Ryujinx.Graphics.Vulkan/HelperShader.cs

@@ -16,6 +16,7 @@ namespace Ryujinx.Graphics.Vulkan
         private readonly IProgram _programColorBlit;
         private readonly IProgram _programColorBlitClearAlpha;
         private readonly IProgram _programColorClear;
+        private readonly IProgram _programStrideChange;
 
         public HelperShader(VulkanRenderer gd, Device device)
         {
@@ -39,14 +40,14 @@ namespace Ryujinx.Graphics.Vulkan
 
             _programColorBlit = gd.CreateProgramWithMinimalLayout(new[]
             {
-                new ShaderSource(ShaderBinaries.ColorBlitVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Glsl),
-                new ShaderSource(ShaderBinaries.ColorBlitFragmentShaderSource, fragmentBindings, ShaderStage.Fragment, TargetLanguage.Glsl),
+                new ShaderSource(ShaderBinaries.ColorBlitVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Spirv),
+                new ShaderSource(ShaderBinaries.ColorBlitFragmentShaderSource, fragmentBindings, ShaderStage.Fragment, TargetLanguage.Spirv),
             });
 
             _programColorBlitClearAlpha = gd.CreateProgramWithMinimalLayout(new[]
             {
-                new ShaderSource(ShaderBinaries.ColorBlitVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Glsl),
-                new ShaderSource(ShaderBinaries.ColorBlitClearAlphaFragmentShaderSource, fragmentBindings, ShaderStage.Fragment, TargetLanguage.Glsl),
+                new ShaderSource(ShaderBinaries.ColorBlitVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Spirv),
+                new ShaderSource(ShaderBinaries.ColorBlitClearAlphaFragmentShaderSource, fragmentBindings, ShaderStage.Fragment, TargetLanguage.Spirv),
             });
 
             var fragmentBindings2 = new ShaderBindings(
@@ -57,8 +58,19 @@ namespace Ryujinx.Graphics.Vulkan
 
             _programColorClear = gd.CreateProgramWithMinimalLayout(new[]
             {
-                new ShaderSource(ShaderBinaries.ColorClearVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Glsl),
-                new ShaderSource(ShaderBinaries.ColorClearFragmentShaderSource, fragmentBindings2, ShaderStage.Fragment, TargetLanguage.Glsl),
+                new ShaderSource(ShaderBinaries.ColorClearVertexShaderSource, vertexBindings, ShaderStage.Vertex, TargetLanguage.Spirv),
+                new ShaderSource(ShaderBinaries.ColorClearFragmentShaderSource, fragmentBindings2, ShaderStage.Fragment, TargetLanguage.Spirv),
+            });
+
+            var strideChangeBindings = new ShaderBindings(
+                new[] { 0 },
+                new[] { 1, 2 },
+                Array.Empty<int>(),
+                Array.Empty<int>());
+
+            _programStrideChange = gd.CreateProgramWithMinimalLayout(new[]
+            {
+                new ShaderSource(ShaderBinaries.ChangeBufferStrideShaderSource, strideChangeBindings, ShaderStage.Compute, TargetLanguage.Spirv),
             });
         }
 
@@ -163,7 +175,7 @@ namespace Ryujinx.Graphics.Vulkan
             _pipeline.SetViewports(viewports, false);
             _pipeline.SetPrimitiveTopology(GAL.PrimitiveTopology.TriangleStrip);
             _pipeline.Draw(4, 1, 0, 0);
-            _pipeline.Finish();
+            _pipeline.Finish(gd, cbs);
 
             gd.BufferManager.Delete(bufferHandle);
         }
@@ -291,45 +303,100 @@ namespace Ryujinx.Graphics.Vulkan
 
         public unsafe void ConvertI8ToI16(VulkanRenderer gd, CommandBufferScoped cbs, BufferHolder src, BufferHolder dst, int srcOffset, int size)
         {
-            // TODO: Do this with a compute shader?
-            var srcBuffer = src.GetBuffer().Get(cbs, srcOffset, size).Value;
-            var dstBuffer = dst.GetBuffer().Get(cbs, 0, size * 2).Value;
-
-            gd.Api.CmdFillBuffer(cbs.CommandBuffer, dstBuffer, 0, Vk.WholeSize, 0);
-
-            var bufferCopy = new BufferCopy[size];
-
-            for (ulong i = 0; i < (ulong)size; i++)
-            {
-                bufferCopy[i] = new BufferCopy((ulong)srcOffset + i, i * 2, 1);
+            ChangeStride(gd, cbs, src, dst, srcOffset, size, 1, 2);
         }
 
+        public unsafe void ChangeStride(VulkanRenderer gd, CommandBufferScoped cbs, BufferHolder src, BufferHolder dst, int srcOffset, int size, int stride, int newStride)
+        {
+            bool supportsUint8 = gd.Capabilities.SupportsShaderInt8;
+
+            int elems = size / stride;
+            int newSize = elems * newStride;
+
+            var srcBufferAuto = src.GetBuffer();
+            var dstBufferAuto = dst.GetBuffer();
+
+            var srcBuffer = srcBufferAuto.Get(cbs, srcOffset, size).Value;
+            var dstBuffer = dstBufferAuto.Get(cbs, 0, newSize).Value;
+
+            var access = supportsUint8 ? AccessFlags.AccessShaderWriteBit : AccessFlags.AccessTransferWriteBit;
+            var stage = supportsUint8 ? PipelineStageFlags.PipelineStageComputeShaderBit : PipelineStageFlags.PipelineStageTransferBit;
+
             BufferHolder.InsertBufferBarrier(
                 gd,
                 cbs.CommandBuffer,
                 dstBuffer,
                 BufferHolder.DefaultAccessFlags,
-                AccessFlags.AccessTransferWriteBit,
+                access,
                 PipelineStageFlags.PipelineStageAllCommandsBit,
-                PipelineStageFlags.PipelineStageTransferBit,
+                stage,
                 0,
-                size * 2);
+                newSize);
+
+            if (supportsUint8)
+            {
+                const int ParamsBufferSize = 16;
+
+                Span<int> shaderParams = stackalloc int[ParamsBufferSize / sizeof(int)];
+
+                shaderParams[0] = stride;
+                shaderParams[1] = newStride;
+                shaderParams[2] = size;
+                shaderParams[3] = srcOffset;
+
+                var bufferHandle = gd.BufferManager.CreateWithHandle(gd, ParamsBufferSize, false);
+
+                gd.BufferManager.SetData<int>(bufferHandle, 0, shaderParams);
+
+                _pipeline.SetCommandBuffer(cbs);
+
+                Span<BufferRange> cbRanges = stackalloc BufferRange[1];
+
+                cbRanges[0] = new BufferRange(bufferHandle, 0, ParamsBufferSize);
+
+                _pipeline.SetUniformBuffers(0, cbRanges);
+
+                Span<Auto<DisposableBuffer>> sbRanges = new Auto<DisposableBuffer>[2];
+
+                sbRanges[0] = srcBufferAuto;
+                sbRanges[1] = dstBufferAuto;
+
+                _pipeline.SetStorageBuffers(1, sbRanges);
+
+                _pipeline.SetProgram(_programStrideChange);
+                _pipeline.DispatchCompute(1, 1, 1);
+
+                gd.BufferManager.Delete(bufferHandle);
+
+                _pipeline.Finish(gd, cbs);
+            }
+            else
+            {
+                gd.Api.CmdFillBuffer(cbs.CommandBuffer, dstBuffer, 0, Vk.WholeSize, 0);
+
+                var bufferCopy = new BufferCopy[elems];
+
+                for (ulong i = 0; i < (ulong)elems; i++)
+                {
+                    bufferCopy[i] = new BufferCopy((ulong)srcOffset + i * (ulong)stride, i * (ulong)newStride, (ulong)stride);
+                }
 
                 fixed (BufferCopy* pBufferCopy = bufferCopy)
                 {
-                gd.Api.CmdCopyBuffer(cbs.CommandBuffer, srcBuffer, dstBuffer, (uint)size, pBufferCopy);
+                    gd.Api.CmdCopyBuffer(cbs.CommandBuffer, srcBuffer, dstBuffer, (uint)elems, pBufferCopy);
+                }
             }
 
             BufferHolder.InsertBufferBarrier(
                 gd,
                 cbs.CommandBuffer,
                 dstBuffer,
-                AccessFlags.AccessTransferWriteBit,
+                access,
                 BufferHolder.DefaultAccessFlags,
-                PipelineStageFlags.PipelineStageTransferBit,
+                stage,
                 PipelineStageFlags.PipelineStageAllCommandsBit,
                 0,
-                size * 2);
+                newSize);
         }
 
         protected virtual void Dispose(bool disposing)

Ryujinx.Graphics.Vulkan/PipelineBase.cs

@@ -2,6 +2,7 @@
 using Ryujinx.Graphics.Shader;
 using Silk.NET.Vulkan;
 using System;
+using System.Numerics;
 
 namespace Ryujinx.Graphics.Vulkan
 {
@@ -50,14 +51,14 @@ namespace Ryujinx.Graphics.Vulkan
 
         private BufferState _indexBuffer;
         private readonly BufferState[] _transformFeedbackBuffers;
-        private readonly BufferState[] _vertexBuffers;
+        private readonly VertexBufferState[] _vertexBuffers;
+        private ulong _vertexBuffersDirty;
         protected Rectangle<int> ClearScissor;
 
         public SupportBufferUpdater SupportBufferUpdater;
 
         private bool _needsIndexBufferRebind;
         private bool _needsTransformFeedbackBuffersRebind;
-        private bool _needsVertexBuffersRebind;
 
         private bool _tfEnabled;
         private bool _tfActive;
@@ -79,14 +80,14 @@ namespace Ryujinx.Graphics.Vulkan
             _descriptorSetUpdater = new DescriptorSetUpdater(gd, this);
 
             _transformFeedbackBuffers = new BufferState[Constants.MaxTransformFeedbackBuffers];
-            _vertexBuffers = new BufferState[Constants.MaxVertexBuffers + 1];
+            _vertexBuffers = new VertexBufferState[Constants.MaxVertexBuffers + 1];
 
             const int EmptyVbSize = 16;
 
             using var emptyVb = gd.BufferManager.Create(gd, EmptyVbSize);
             emptyVb.SetData(0, new byte[EmptyVbSize]);
-            _vertexBuffers[0] = new BufferState(emptyVb.GetBuffer(), 0, EmptyVbSize, 0UL);
-            _needsVertexBuffersRebind = true;
+            _vertexBuffers[0] = new VertexBufferState(emptyVb.GetBuffer(), 0, EmptyVbSize, 0);
+            _vertexBuffersDirty = ulong.MaxValue >> (64 - _vertexBuffers.Length);
 
             ClearScissor = new Rectangle<int>(0, 0, 0xffff, 0xffff);
 
@@ -229,6 +230,17 @@ namespace Ryujinx.Graphics.Vulkan
             BufferHolder.Copy(Gd, Cbs, src, dst, srcOffset, dstOffset, size);
         }
 
+        public void DirtyVertexBuffer(Auto<DisposableBuffer> buffer)
+        {
+            for (int i = 0; i < _vertexBuffers.Length; i++)
+            {
+                if (_vertexBuffers[i].BoundEquals(buffer))
+                {
+                    _vertexBuffersDirty |= 1UL << i;
+                }
+            }
+        }
+
         public void DispatchCompute(int groupsX, int groupsY, int groupsZ)
         {
             if (!_program.IsLinked)
@@ -345,6 +357,11 @@ namespace Ryujinx.Graphics.Vulkan
             _tfEnabled = false;
         }
 
+        public bool IsCommandBufferActive(CommandBuffer cb)
+        {
+            return CommandBuffer.Handle == cb.Handle;
+        }
+
         public void MultiDrawIndirectCount(BufferRange indirectBuffer, BufferRange parameterBuffer, int maxDrawCount, int stride)
         {
             if (!Gd.Capabilities.SupportsIndirectParameters)
@@ -689,6 +706,11 @@ namespace Ryujinx.Graphics.Vulkan
             _descriptorSetUpdater.SetStorageBuffers(CommandBuffer, first, buffers);
         }
 
+        public void SetStorageBuffers(int first, ReadOnlySpan<Auto<DisposableBuffer>> buffers)
+        {
+            _descriptorSetUpdater.SetStorageBuffers(CommandBuffer, first, buffers);
+        }
+
         public void SetTextureAndSampler(ShaderStage stage, int binding, ITexture texture, ISampler sampler)
         {
             _descriptorSetUpdater.SetTextureAndSampler(Cbs, stage, binding, texture, sampler);
@@ -732,12 +754,22 @@ namespace Ryujinx.Graphics.Vulkan
         {
             var formatCapabilities = Gd.FormatCapabilities;
 
+            Span<int> newVbScalarSizes = stackalloc int[Constants.MaxVertexBuffers];
+
             int count = Math.Min(Constants.MaxVertexAttributes, vertexAttribs.Length);
+            uint dirtyVbSizes = 0;
 
             for (int i = 0; i < count; i++)
             {
                 var attribute = vertexAttribs[i];
-                var bufferIndex = attribute.IsZero ? 0 : attribute.BufferIndex + 1;
+                var rawIndex = attribute.BufferIndex;
+                var bufferIndex = attribute.IsZero ? 0 : rawIndex + 1;
+
+                if (!attribute.IsZero)
+                {
+                    newVbScalarSizes[rawIndex] = Math.Max(newVbScalarSizes[rawIndex], attribute.Format.GetScalarSize());
+                    dirtyVbSizes |= 1u << rawIndex;
+                }
 
                 _newState.Internal.VertexAttributeDescriptions[i] = new VertexInputAttributeDescription(
                     (uint)i,
@@ -746,6 +778,21 @@ namespace Ryujinx.Graphics.Vulkan
                     (uint)attribute.Offset);
             }
 
+            while (dirtyVbSizes != 0)
+            {
+                int dirtyBit = BitOperations.TrailingZeroCount(dirtyVbSizes);
+
+                ref var buffer = ref _vertexBuffers[dirtyBit + 1];
+
+                if (buffer.AttributeScalarAlignment != newVbScalarSizes[dirtyBit])
+                {
+                    _vertexBuffersDirty |= 1UL << (dirtyBit + 1);
+                    buffer.AttributeScalarAlignment = newVbScalarSizes[dirtyBit];
+                }
+
+                dirtyVbSizes &= ~(1u << dirtyBit);
+            }
+
             _newState.VertexAttributeDescriptionsCount = (uint)count;
             SignalStateChange();
         }
@@ -792,14 +839,37 @@ namespace Ryujinx.Graphics.Vulkan
                             }
                         }
 
-                        _vertexBuffers[binding].Dispose();
-                        _vertexBuffers[binding] = new BufferState(
+                        ref var buffer = ref _vertexBuffers[binding];
+                        int oldScalarAlign = buffer.AttributeScalarAlignment;
+
+                        buffer.Dispose();
+
+                        if ((vertexBuffer.Stride % FormatExtensions.MaxBufferFormatScalarSize) == 0)
+                        {
+                            buffer = new VertexBufferState(
                                 vb,
+                                descriptorIndex,
                                 vertexBuffer.Buffer.Offset,
                                 vbSize,
-                            (ulong)vertexBuffer.Stride);
+                                vertexBuffer.Stride);
 
-                        _vertexBuffers[binding].BindVertexBuffer(Gd, Cbs, (uint)binding);
+                            buffer.BindVertexBuffer(Gd, Cbs, (uint)binding, ref _newState);
+                        }
+                        else
+                        {
+                            // May need to be rewritten. Bind this buffer before draw.
+
+                            buffer = new VertexBufferState(
+                                vertexBuffer.Buffer.Handle,
+                                descriptorIndex,
+                                vertexBuffer.Buffer.Offset,
+                                vbSize,
+                                vertexBuffer.Stride);
+
+                            _vertexBuffersDirty |= 1UL << binding;
+                        }
+
+                        buffer.AttributeScalarAlignment = oldScalarAlign;
                     }
                 }
             }
@@ -907,7 +977,7 @@ namespace Ryujinx.Graphics.Vulkan
         {
             _needsIndexBufferRebind = true;
             _needsTransformFeedbackBuffersRebind = true;
-            _needsVertexBuffersRebind = true;
+            _vertexBuffersDirty = ulong.MaxValue >> (64 - _vertexBuffers.Length);
 
             _descriptorSetUpdater.SignalCommandBufferChange();
             _dynamicState.ForceAllDirty();
@@ -1053,13 +1123,6 @@ namespace Ryujinx.Graphics.Vulkan
             // Commit changes to the support buffer before drawing.
             SupportBufferUpdater.Commit();
 
-            if (_stateDirty || Pbp != pbp)
-            {
-                CreatePipeline(pbp);
-                _stateDirty = false;
-                Pbp = pbp;
-            }
-
             if (_needsIndexBufferRebind)
             {
                 _indexBuffer.BindIndexBuffer(Gd.Api, Cbs);
@@ -1078,14 +1141,23 @@ namespace Ryujinx.Graphics.Vulkan
                 _needsTransformFeedbackBuffersRebind = false;
             }
 
-            if (_needsVertexBuffersRebind)
+            if (_vertexBuffersDirty != 0)
             {
-                for (int i = 0; i < Constants.MaxVertexBuffers + 1; i++)
+                while (_vertexBuffersDirty != 0)
                 {
-                    _vertexBuffers[i].BindVertexBuffer(Gd, Cbs, (uint)i);
+                    int i = BitOperations.TrailingZeroCount(_vertexBuffersDirty);
+
+                    _vertexBuffers[i].BindVertexBuffer(Gd, Cbs, (uint)i, ref _newState);
+
+                    _vertexBuffersDirty &= ~(1u << i);
+                }
             }
 
-                _needsVertexBuffersRebind = false;
+            if (_stateDirty || Pbp != pbp)
+            {
+                CreatePipeline(pbp);
+                _stateDirty = false;
+                Pbp = pbp;
             }
 
             _descriptorSetUpdater.UpdateAndBindDescriptorSets(Cbs, pbp);

Ryujinx.Graphics.Vulkan/PipelineConverter.cs

@@ -202,6 +202,9 @@ namespace Ryujinx.Graphics.Vulkan
             pipeline.Topology = state.Topology.Convert();
 
             int vaCount = Math.Min(Constants.MaxVertexAttributes, state.VertexAttribCount);
+            int vbCount = Math.Min(Constants.MaxVertexBuffers, state.VertexBufferCount);
+
+            Span<int> vbScalarSizes = stackalloc int[vbCount];
 
             for (int i = 0; i < vaCount; i++)
             {
@@ -213,13 +216,16 @@ namespace Ryujinx.Graphics.Vulkan
                     (uint)bufferIndex,
                     gd.FormatCapabilities.ConvertToVertexVkFormat(attribute.Format),
                     (uint)attribute.Offset);
+
+                if (!attribute.IsZero && bufferIndex < vbCount)
+                {
+                    vbScalarSizes[bufferIndex - 1] = Math.Max(attribute.Format.GetScalarSize(), vbScalarSizes[bufferIndex - 1]);
+                }
             }
 
             int descriptorIndex = 1;
             pipeline.Internal.VertexBindingDescriptions[0] = new VertexInputBindingDescription(0, 0, VertexInputRate.Vertex);
 
-            int vbCount = Math.Min(Constants.MaxVertexBuffers, state.VertexBufferCount);
-
             for (int i = 0; i < vbCount; i++)
             {
                 var vertexBuffer = state.VertexBuffers[i];
@@ -228,10 +234,17 @@ namespace Ryujinx.Graphics.Vulkan
                 {
                     var inputRate = vertexBuffer.Divisor != 0 ? VertexInputRate.Instance : VertexInputRate.Vertex;
 
+                    int alignedStride = vertexBuffer.Stride;
+
+                    if (gd.NeedsVertexBufferAlignment(vbScalarSizes[i], out int alignment))
+                    {
+                        alignedStride = (vertexBuffer.Stride + (alignment - 1)) & -alignment;
+                    }
+
                     // TODO: Support divisor > 1
                     pipeline.Internal.VertexBindingDescriptions[descriptorIndex++] = new VertexInputBindingDescription(
                         (uint)i + 1,
-                        (uint)vertexBuffer.Stride,
+                        (uint)alignedStride,
                         inputRate);
                 }
             }

Ryujinx.Graphics.Vulkan/PipelineFull.cs

@@ -199,6 +199,16 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        public void Restore()
+        {
+            if (Pipeline != null)
+            {
+                Gd.Api.CmdBindPipeline(CommandBuffer, Pbp, Pipeline.Get(Cbs).Value);
+            }
+
+            SignalCommandBufferChange();
+        }
+
         public void FlushCommandsImpl()
         {
             EndRenderPass();
@@ -220,18 +230,13 @@ namespace Ryujinx.Graphics.Vulkan
 
             // Restore per-command buffer state.
 
-            if (Pipeline != null)
-            {
-                Gd.Api.CmdBindPipeline(CommandBuffer, Pbp, Pipeline.Get(Cbs).Value);
-            }
-
             foreach (var queryPool in _activeQueries)
             {
                 Gd.Api.CmdResetQueryPool(CommandBuffer, queryPool, 0, 1);
                 Gd.Api.CmdBeginQuery(CommandBuffer, queryPool, 0, 0);
             }
 
-            SignalCommandBufferChange();
+            Restore();
         }
 
         public void BeginQuery(BufferedQuery query, QueryPool pool, bool needsReset)

Ryujinx.Graphics.Vulkan/PipelineHelperShader.cs

@@ -40,5 +40,15 @@ namespace Ryujinx.Graphics.Vulkan
         {
             EndRenderPass();
         }
+
+        public void Finish(VulkanRenderer gd, CommandBufferScoped cbs)
+        {
+            Finish();
+
+            if (gd.PipelineInternal.IsCommandBufferActive(cbs.CommandBuffer))
+            {
+                gd.PipelineInternal.Restore();
+            }
+        }
     }
 }

Ryujinx.Graphics.Vulkan/PipelineLayoutFactory.cs

@@ -142,18 +142,20 @@ namespace Ryujinx.Graphics.Vulkan
             int stagesCount = shaders.Length;
 
             int uCount = 0;
+            int sCount = 0;
             int tCount = 0;
             int iCount = 0;
 
             foreach (var shader in shaders)
             {
                 uCount += shader.Bindings.UniformBufferBindings.Count;
+                sCount += shader.Bindings.StorageBufferBindings.Count;
                 tCount += shader.Bindings.TextureBindings.Count;
                 iCount += shader.Bindings.ImageBindings.Count;
             }
 
             DescriptorSetLayoutBinding* uLayoutBindings = stackalloc DescriptorSetLayoutBinding[uCount];
-            DescriptorSetLayoutBinding* sLayoutBindings = stackalloc DescriptorSetLayoutBinding[stagesCount];
+            DescriptorSetLayoutBinding* sLayoutBindings = stackalloc DescriptorSetLayoutBinding[sCount];
             DescriptorSetLayoutBinding* tLayoutBindings = stackalloc DescriptorSetLayoutBinding[tCount];
             DescriptorSetLayoutBinding* iLayoutBindings = stackalloc DescriptorSetLayoutBinding[iCount];
 
@@ -180,22 +182,11 @@ namespace Ryujinx.Graphics.Vulkan
                     }
                 }
 
-                void SetStorage(DescriptorSetLayoutBinding* bindings, ref int start, int count)
-                {
-                    bindings[start++] = new DescriptorSetLayoutBinding
-                    {
-                        Binding = (uint)start,
-                        DescriptorType = DescriptorType.StorageBuffer,
-                        DescriptorCount = (uint)count,
-                        StageFlags = stageFlags
-                    };
-                }
-
                 // TODO: Support buffer textures and images here.
                 // This is only used for the helper shaders on the backend, and we don't use buffer textures on them
                 // so far, so it's not really necessary right now.
                 Set(uLayoutBindings, DescriptorType.UniformBuffer, ref uIndex, shader.Bindings.UniformBufferBindings);
-                SetStorage(sLayoutBindings, ref sIndex, shader.Bindings.StorageBufferBindings.Count);
+                Set(sLayoutBindings, DescriptorType.StorageBuffer, ref sIndex, shader.Bindings.StorageBufferBindings);
                 Set(tLayoutBindings, DescriptorType.CombinedImageSampler, ref tIndex, shader.Bindings.TextureBindings);
                 Set(iLayoutBindings, DescriptorType.StorageImage, ref iIndex, shader.Bindings.ImageBindings);
             }
@@ -213,7 +204,7 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 SType = StructureType.DescriptorSetLayoutCreateInfo,
                 PBindings = sLayoutBindings,
-                BindingCount = (uint)stagesCount
+                BindingCount = (uint)sCount
             };
 
             var tDescriptorSetLayoutCreateInfo = new DescriptorSetLayoutCreateInfo()

Ryujinx.Graphics.Vulkan/Shaders/ChangeBufferStrideShaderSource.comp

@@ -0,0 +1,64 @@
+#version 450 core
+
+#extension GL_EXT_shader_8bit_storage : require
+
+layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (std140, set = 0, binding = 0) uniform stride_arguments
+{
+    ivec4 stride_arguments_data;
+};
+
+layout (std430, set = 1, binding = 1) buffer in_s
+{
+    uint8_t[] in_data;
+};
+
+layout (std430, set = 1, binding = 2) buffer out_s
+{
+    uint8_t[] out_data;
+};
+
+void main()
+{
+    // Determine what slice of the stride copies this invocation will perform.
+
+    int sourceStride = stride_arguments_data.x;
+    int targetStride = stride_arguments_data.y;
+    int bufferSize = stride_arguments_data.z;
+    int sourceOffset = stride_arguments_data.w;
+
+    int strideRemainder = targetStride - sourceStride;
+    int invocations = int(gl_WorkGroupSize.x);
+
+    int copiesRequired = bufferSize / sourceStride;
+
+    // Find the copies that this invocation should perform.
+    
+    // - Copies that all invocations perform.
+    int allInvocationCopies = copiesRequired / invocations;
+
+    // - Extra remainder copy that this invocation performs.
+    int index = int(gl_LocalInvocationID.x);
+    int extra = (index < (copiesRequired % invocations)) ? 1 : 0;
+
+    int copyCount = allInvocationCopies + extra;
+
+    // Finally, get the starting offset. Make sure to count extra copies.
+
+    int startCopy = allInvocationCopies * index + min(copiesRequired % invocations, index);
+
+    int srcOffset = sourceOffset + startCopy * sourceStride;
+    int dstOffset = startCopy * targetStride;
+
+    // Perform the copies for this region
+    for (int i=0; i<copyCount; i++) {
+        for (int j=0; j<sourceStride; j++) {
+            out_data[dstOffset++] = in_data[srcOffset++];
+        }
+
+        for (int j=0; j<strideRemainder; j++) {
+            out_data[dstOffset++] = uint8_t(0);
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/Shaders/ShaderBinaries.cs

@@ -2,6 +2,249 @@ namespace Ryujinx.Graphics.Vulkan.Shaders
 {
     static class ShaderBinaries
     {
+        public static readonly byte[] ChangeBufferStrideShaderSource = new byte[]
+        {
+            0x03, 0x02, 0x23, 0x07, 0x00, 0x05, 0x01, 0x00, 0x0A, 0x00, 0x0D, 0x00, 0x8E, 0x00, 0x00, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x01, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00,
+            0x60, 0x11, 0x00, 0x00, 0x0B, 0x00, 0x06, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x4C, 0x53, 0x4C,
+            0x2E, 0x73, 0x74, 0x64, 0x2E, 0x34, 0x35, 0x30, 0x00, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x03, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x09, 0x00, 0x05, 0x00, 0x00, 0x00,
+            0x04, 0x00, 0x00, 0x00, 0x6D, 0x61, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00,
+            0x30, 0x00, 0x00, 0x00, 0x6C, 0x00, 0x00, 0x00, 0x72, 0x00, 0x00, 0x00, 0x10, 0x00, 0x06, 0x00,
+            0x04, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+            0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x03, 0x00, 0x02, 0x00, 0x00, 0x00, 0xC2, 0x01, 0x00, 0x00,
+            0x04, 0x00, 0x08, 0x00, 0x47, 0x4C, 0x5F, 0x45, 0x58, 0x54, 0x5F, 0x73, 0x68, 0x61, 0x64, 0x65,
+            0x72, 0x5F, 0x38, 0x62, 0x69, 0x74, 0x5F, 0x73, 0x74, 0x6F, 0x72, 0x61, 0x67, 0x65, 0x00, 0x00,
+            0x04, 0x00, 0x0A, 0x00, 0x47, 0x4C, 0x5F, 0x47, 0x4F, 0x4F, 0x47, 0x4C, 0x45, 0x5F, 0x63, 0x70,
+            0x70, 0x5F, 0x73, 0x74, 0x79, 0x6C, 0x65, 0x5F, 0x6C, 0x69, 0x6E, 0x65, 0x5F, 0x64, 0x69, 0x72,
+            0x65, 0x63, 0x74, 0x69, 0x76, 0x65, 0x00, 0x00, 0x04, 0x00, 0x08, 0x00, 0x47, 0x4C, 0x5F, 0x47,
+            0x4F, 0x4F, 0x47, 0x4C, 0x45, 0x5F, 0x69, 0x6E, 0x63, 0x6C, 0x75, 0x64, 0x65, 0x5F, 0x64, 0x69,
+            0x72, 0x65, 0x63, 0x74, 0x69, 0x76, 0x65, 0x00, 0x05, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00,
+            0x6D, 0x61, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x06, 0x00, 0x08, 0x00, 0x00, 0x00,
+            0x73, 0x6F, 0x75, 0x72, 0x63, 0x65, 0x53, 0x74, 0x72, 0x69, 0x64, 0x65, 0x00, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x07, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x73, 0x74, 0x72, 0x69, 0x64, 0x65, 0x5F, 0x61,
+            0x72, 0x67, 0x75, 0x6D, 0x65, 0x6E, 0x74, 0x73, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x09, 0x00,
+            0x0A, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x73, 0x74, 0x72, 0x69, 0x64, 0x65, 0x5F, 0x61,
+            0x72, 0x67, 0x75, 0x6D, 0x65, 0x6E, 0x74, 0x73, 0x5F, 0x64, 0x61, 0x74, 0x61, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x03, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x06, 0x00,
+            0x13, 0x00, 0x00, 0x00, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x53, 0x74, 0x72, 0x69, 0x64, 0x65,
+            0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x05, 0x00, 0x17, 0x00, 0x00, 0x00, 0x62, 0x75, 0x66, 0x66,
+            0x65, 0x72, 0x53, 0x69, 0x7A, 0x65, 0x00, 0x00, 0x05, 0x00, 0x06, 0x00, 0x1B, 0x00, 0x00, 0x00,
+            0x73, 0x6F, 0x75, 0x72, 0x63, 0x65, 0x4F, 0x66, 0x66, 0x73, 0x65, 0x74, 0x00, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x06, 0x00, 0x1F, 0x00, 0x00, 0x00, 0x73, 0x74, 0x72, 0x69, 0x64, 0x65, 0x52, 0x65,
+            0x6D, 0x61, 0x69, 0x6E, 0x64, 0x65, 0x72, 0x00, 0x05, 0x00, 0x05, 0x00, 0x23, 0x00, 0x00, 0x00,
+            0x69, 0x6E, 0x76, 0x6F, 0x63, 0x61, 0x74, 0x69, 0x6F, 0x6E, 0x73, 0x00, 0x05, 0x00, 0x06, 0x00,
+            0x25, 0x00, 0x00, 0x00, 0x63, 0x6F, 0x70, 0x69, 0x65, 0x73, 0x52, 0x65, 0x71, 0x75, 0x69, 0x72,
+            0x65, 0x64, 0x00, 0x00, 0x05, 0x00, 0x07, 0x00, 0x29, 0x00, 0x00, 0x00, 0x61, 0x6C, 0x6C, 0x49,
+            0x6E, 0x76, 0x6F, 0x63, 0x61, 0x74, 0x69, 0x6F, 0x6E, 0x43, 0x6F, 0x70, 0x69, 0x65, 0x73, 0x00,
+            0x05, 0x00, 0x04, 0x00, 0x2D, 0x00, 0x00, 0x00, 0x69, 0x6E, 0x64, 0x65, 0x78, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x08, 0x00, 0x30, 0x00, 0x00, 0x00, 0x67, 0x6C, 0x5F, 0x4C, 0x6F, 0x63, 0x61, 0x6C,
+            0x49, 0x6E, 0x76, 0x6F, 0x63, 0x61, 0x74, 0x69, 0x6F, 0x6E, 0x49, 0x44, 0x00, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x04, 0x00, 0x35, 0x00, 0x00, 0x00, 0x65, 0x78, 0x74, 0x72, 0x61, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x05, 0x00, 0x3E, 0x00, 0x00, 0x00, 0x63, 0x6F, 0x70, 0x79, 0x43, 0x6F, 0x75, 0x6E,
+            0x74, 0x00, 0x00, 0x00, 0x05, 0x00, 0x05, 0x00, 0x42, 0x00, 0x00, 0x00, 0x73, 0x74, 0x61, 0x72,
+            0x74, 0x43, 0x6F, 0x70, 0x79, 0x00, 0x00, 0x00, 0x05, 0x00, 0x05, 0x00, 0x4C, 0x00, 0x00, 0x00,
+            0x73, 0x72, 0x63, 0x4F, 0x66, 0x66, 0x73, 0x65, 0x74, 0x00, 0x00, 0x00, 0x05, 0x00, 0x05, 0x00,
+            0x52, 0x00, 0x00, 0x00, 0x64, 0x73, 0x74, 0x4F, 0x66, 0x66, 0x73, 0x65, 0x74, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x03, 0x00, 0x56, 0x00, 0x00, 0x00, 0x69, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00,
+            0x5F, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x00, 0x00, 0x05, 0x00, 0x04, 0x00, 0x6A, 0x00, 0x00, 0x00,
+            0x6F, 0x75, 0x74, 0x5F, 0x73, 0x00, 0x00, 0x00, 0x06, 0x00, 0x06, 0x00, 0x6A, 0x00, 0x00, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x6F, 0x75, 0x74, 0x5F, 0x64, 0x61, 0x74, 0x61, 0x00, 0x00, 0x00, 0x00,
+            0x05, 0x00, 0x03, 0x00, 0x6C, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x04, 0x00,
+            0x70, 0x00, 0x00, 0x00, 0x69, 0x6E, 0x5F, 0x73, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x05, 0x00,
+            0x70, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x69, 0x6E, 0x5F, 0x64, 0x61, 0x74, 0x61, 0x00,
+            0x05, 0x00, 0x03, 0x00, 0x72, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00,
+            0x7B, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x0A, 0x00, 0x00, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00,
+            0x0A, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x0C, 0x00, 0x00, 0x00,
+            0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x0C, 0x00, 0x00, 0x00,
+            0x21, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x30, 0x00, 0x00, 0x00,
+            0x0B, 0x00, 0x00, 0x00, 0x1B, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x69, 0x00, 0x00, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x6A, 0x00, 0x00, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00,
+            0x6A, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x6C, 0x00, 0x00, 0x00,
+            0x22, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x6C, 0x00, 0x00, 0x00,
+            0x21, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x6F, 0x00, 0x00, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x70, 0x00, 0x00, 0x00,
+            0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00,
+            0x70, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x72, 0x00, 0x00, 0x00,
+            0x22, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x72, 0x00, 0x00, 0x00,
+            0x21, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x8D, 0x00, 0x00, 0x00,
+            0x0B, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x13, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00,
+            0x21, 0x00, 0x03, 0x00, 0x03, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00,
+            0x09, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x1E, 0x00, 0x03, 0x00,
+            0x0A, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00,
+            0x02, 0x00, 0x00, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00,
+            0x0C, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x0D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00,
+            0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00,
+            0x0F, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00,
+            0x02, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00,
+            0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00,
+            0x18, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00,
+            0x1C, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x24, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x2E, 0x00, 0x00, 0x00,
+            0x0E, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x2F, 0x00, 0x00, 0x00,
+            0x01, 0x00, 0x00, 0x00, 0x2E, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x2F, 0x00, 0x00, 0x00,
+            0x30, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x31, 0x00, 0x00, 0x00,
+            0x01, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x14, 0x00, 0x02, 0x00, 0x3A, 0x00, 0x00, 0x00,
+            0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+            0x15, 0x00, 0x04, 0x00, 0x68, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+            0x1D, 0x00, 0x03, 0x00, 0x69, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0x1E, 0x00, 0x03, 0x00,
+            0x6A, 0x00, 0x00, 0x00, 0x69, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x6B, 0x00, 0x00, 0x00,
+            0x0C, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x6B, 0x00, 0x00, 0x00,
+            0x6C, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x1D, 0x00, 0x03, 0x00, 0x6F, 0x00, 0x00, 0x00,
+            0x68, 0x00, 0x00, 0x00, 0x1E, 0x00, 0x03, 0x00, 0x70, 0x00, 0x00, 0x00, 0x6F, 0x00, 0x00, 0x00,
+            0x20, 0x00, 0x04, 0x00, 0x71, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x70, 0x00, 0x00, 0x00,
+            0x3B, 0x00, 0x04, 0x00, 0x71, 0x00, 0x00, 0x00, 0x72, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00,
+            0x20, 0x00, 0x04, 0x00, 0x75, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00,
+            0x2B, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x8C, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00,
+            0x2C, 0x00, 0x06, 0x00, 0x2E, 0x00, 0x00, 0x00, 0x8D, 0x00, 0x00, 0x00, 0x8C, 0x00, 0x00, 0x00,
+            0x14, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x36, 0x00, 0x05, 0x00, 0x02, 0x00, 0x00, 0x00,
+            0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00,
+            0x05, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x1B, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x1F, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x4C, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x56, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x7B, 0x00, 0x00, 0x00,
+            0x07, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x10, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00,
+            0x0C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00,
+            0x08, 0x00, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x10, 0x00, 0x00, 0x00,
+            0x15, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x13, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00,
+            0x10, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00,
+            0x18, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x1A, 0x00, 0x00, 0x00,
+            0x19, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x17, 0x00, 0x00, 0x00, 0x1A, 0x00, 0x00, 0x00,
+            0x41, 0x00, 0x06, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1D, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00,
+            0x0D, 0x00, 0x00, 0x00, 0x1C, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x1E, 0x00, 0x00, 0x00, 0x1D, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x1B, 0x00, 0x00, 0x00,
+            0x1E, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00,
+            0x13, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00,
+            0x08, 0x00, 0x00, 0x00, 0x82, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00,
+            0x20, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x1F, 0x00, 0x00, 0x00,
+            0x22, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x23, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x26, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
+            0x87, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x26, 0x00, 0x00, 0x00,
+            0x27, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x25, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2A, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00,
+            0x87, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x00, 0x00, 0x2A, 0x00, 0x00, 0x00,
+            0x2B, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x29, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x00, 0x00,
+            0x41, 0x00, 0x05, 0x00, 0x31, 0x00, 0x00, 0x00, 0x32, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00,
+            0x0F, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x33, 0x00, 0x00, 0x00,
+            0x32, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00,
+            0x33, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x2D, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x37, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00,
+            0x8B, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x39, 0x00, 0x00, 0x00, 0x37, 0x00, 0x00, 0x00,
+            0x38, 0x00, 0x00, 0x00, 0xB1, 0x00, 0x05, 0x00, 0x3A, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x00, 0x00,
+            0x36, 0x00, 0x00, 0x00, 0x39, 0x00, 0x00, 0x00, 0xA9, 0x00, 0x06, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x35, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x3E, 0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x43, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00, 0x84, 0x00, 0x05, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x45, 0x00, 0x00, 0x00, 0x43, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x46, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00,
+            0x8B, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x46, 0x00, 0x00, 0x00,
+            0x47, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x49, 0x00, 0x00, 0x00,
+            0x2D, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4A, 0x00, 0x00, 0x00,
+            0x01, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x49, 0x00, 0x00, 0x00,
+            0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4B, 0x00, 0x00, 0x00, 0x45, 0x00, 0x00, 0x00,
+            0x4A, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x42, 0x00, 0x00, 0x00, 0x4B, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4D, 0x00, 0x00, 0x00, 0x1B, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4E, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4F, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
+            0x84, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, 0x4E, 0x00, 0x00, 0x00,
+            0x4F, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x51, 0x00, 0x00, 0x00,
+            0x4D, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x4C, 0x00, 0x00, 0x00,
+            0x51, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00,
+            0x42, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00,
+            0x13, 0x00, 0x00, 0x00, 0x84, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x55, 0x00, 0x00, 0x00,
+            0x53, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x52, 0x00, 0x00, 0x00,
+            0x55, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x56, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00,
+            0xF9, 0x00, 0x02, 0x00, 0x57, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x57, 0x00, 0x00, 0x00,
+            0xF6, 0x00, 0x04, 0x00, 0x59, 0x00, 0x00, 0x00, 0x5A, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+            0xF9, 0x00, 0x02, 0x00, 0x5B, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x5B, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x5C, 0x00, 0x00, 0x00, 0x56, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x00, 0x00,
+            0xB1, 0x00, 0x05, 0x00, 0x3A, 0x00, 0x00, 0x00, 0x5E, 0x00, 0x00, 0x00, 0x5C, 0x00, 0x00, 0x00,
+            0x5D, 0x00, 0x00, 0x00, 0xFA, 0x00, 0x04, 0x00, 0x5E, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00,
+            0x59, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x58, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00,
+            0x5F, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, 0x60, 0x00, 0x00, 0x00,
+            0xF8, 0x00, 0x02, 0x00, 0x60, 0x00, 0x00, 0x00, 0xF6, 0x00, 0x04, 0x00, 0x62, 0x00, 0x00, 0x00,
+            0x63, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, 0x64, 0x00, 0x00, 0x00,
+            0xF8, 0x00, 0x02, 0x00, 0x64, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x65, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x66, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0xB1, 0x00, 0x05, 0x00, 0x3A, 0x00, 0x00, 0x00,
+            0x67, 0x00, 0x00, 0x00, 0x65, 0x00, 0x00, 0x00, 0x66, 0x00, 0x00, 0x00, 0xFA, 0x00, 0x04, 0x00,
+            0x67, 0x00, 0x00, 0x00, 0x61, 0x00, 0x00, 0x00, 0x62, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00,
+            0x61, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x6D, 0x00, 0x00, 0x00,
+            0x52, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x6E, 0x00, 0x00, 0x00,
+            0x6D, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x52, 0x00, 0x00, 0x00,
+            0x6E, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x73, 0x00, 0x00, 0x00,
+            0x4C, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x74, 0x00, 0x00, 0x00,
+            0x73, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x4C, 0x00, 0x00, 0x00,
+            0x74, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x75, 0x00, 0x00, 0x00, 0x76, 0x00, 0x00, 0x00,
+            0x72, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x73, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x68, 0x00, 0x00, 0x00, 0x77, 0x00, 0x00, 0x00, 0x76, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00,
+            0x75, 0x00, 0x00, 0x00, 0x78, 0x00, 0x00, 0x00, 0x6C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00,
+            0x6D, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x78, 0x00, 0x00, 0x00, 0x77, 0x00, 0x00, 0x00,
+            0xF9, 0x00, 0x02, 0x00, 0x63, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x63, 0x00, 0x00, 0x00,
+            0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x79, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x00, 0x00,
+            0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x7A, 0x00, 0x00, 0x00, 0x79, 0x00, 0x00, 0x00,
+            0x3C, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00, 0x5F, 0x00, 0x00, 0x00, 0x7A, 0x00, 0x00, 0x00,
+            0xF9, 0x00, 0x02, 0x00, 0x60, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x62, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x7B, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x7C, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x7C, 0x00, 0x00, 0x00, 0xF6, 0x00, 0x04, 0x00,
+            0x7E, 0x00, 0x00, 0x00, 0x7F, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x80, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x80, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x81, 0x00, 0x00, 0x00, 0x7B, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x82, 0x00, 0x00, 0x00, 0x1F, 0x00, 0x00, 0x00, 0xB1, 0x00, 0x05, 0x00,
+            0x3A, 0x00, 0x00, 0x00, 0x83, 0x00, 0x00, 0x00, 0x81, 0x00, 0x00, 0x00, 0x82, 0x00, 0x00, 0x00,
+            0xFA, 0x00, 0x04, 0x00, 0x83, 0x00, 0x00, 0x00, 0x7D, 0x00, 0x00, 0x00, 0x7E, 0x00, 0x00, 0x00,
+            0xF8, 0x00, 0x02, 0x00, 0x7D, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x84, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00,
+            0x85, 0x00, 0x00, 0x00, 0x84, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x3E, 0x00, 0x03, 0x00,
+            0x52, 0x00, 0x00, 0x00, 0x85, 0x00, 0x00, 0x00, 0x71, 0x00, 0x04, 0x00, 0x68, 0x00, 0x00, 0x00,
+            0x86, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x75, 0x00, 0x00, 0x00,
+            0x87, 0x00, 0x00, 0x00, 0x6C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x84, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x87, 0x00, 0x00, 0x00, 0x86, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x7F, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x7F, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x88, 0x00, 0x00, 0x00, 0x7B, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x89, 0x00, 0x00, 0x00, 0x88, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x7B, 0x00, 0x00, 0x00, 0x89, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x7C, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x7E, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x5A, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x5A, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x8A, 0x00, 0x00, 0x00, 0x56, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00,
+            0x06, 0x00, 0x00, 0x00, 0x8B, 0x00, 0x00, 0x00, 0x8A, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00,
+            0x3E, 0x00, 0x03, 0x00, 0x56, 0x00, 0x00, 0x00, 0x8B, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00,
+            0x57, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x59, 0x00, 0x00, 0x00, 0xFD, 0x00, 0x01, 0x00,
+            0x38, 0x00, 0x01, 0x00
+        };
+
         public static readonly byte[] ColorBlitClearAlphaFragmentShaderSource = new byte[]
         {
             0x03, 0x02, 0x23, 0x07, 0x00, 0x00, 0x01, 0x00, 0x0A, 0x00, 0x08, 0x00, 0x1B, 0x00, 0x00, 0x00,

Ryujinx.Graphics.Vulkan/VertexBufferState.cs

@@ -0,0 +1,131 @@
+using BufferHandle = Ryujinx.Graphics.GAL.BufferHandle;
+
+namespace Ryujinx.Graphics.Vulkan
+{
+    internal struct VertexBufferState
+    {
+        public static VertexBufferState Null => new VertexBufferState(null, 0, 0, 0);
+
+        private readonly int _offset;
+        private readonly int _size;
+        private readonly int _stride;
+
+        private readonly BufferHandle _handle;
+        private Auto<DisposableBuffer> _buffer;
+
+        internal readonly int DescriptorIndex;
+        internal int AttributeScalarAlignment;
+
+        public VertexBufferState(Auto<DisposableBuffer> buffer, int descriptorIndex, int offset, int size, int stride = 0)
+        {
+            _buffer = buffer;
+            _handle = BufferHandle.Null;
+
+            _offset = offset;
+            _size = size;
+            _stride = stride;
+
+            DescriptorIndex = descriptorIndex;
+            AttributeScalarAlignment = 1;
+
+            buffer?.IncrementReferenceCount();
+        }
+
+        public VertexBufferState(BufferHandle handle, int descriptorIndex, int offset, int size, int stride = 0)
+        {
+            // This buffer state may be rewritten at bind time, so it must be retrieved on bind.
+
+            _buffer = null;
+            _handle = handle;
+
+            _offset = offset;
+            _size = size;
+            _stride = stride;
+
+            DescriptorIndex = descriptorIndex;
+            AttributeScalarAlignment = 1;
+        }
+
+        public void BindVertexBuffer(VulkanRenderer gd, CommandBufferScoped cbs, uint binding, ref PipelineState state)
+        {
+            var autoBuffer = _buffer;
+
+            if (_handle != BufferHandle.Null)
+            {
+                // May need to restride the vertex buffer.
+
+                if (gd.NeedsVertexBufferAlignment(AttributeScalarAlignment, out int alignment) && (_stride % alignment) != 0)
+                {
+                    autoBuffer = gd.BufferManager.GetAlignedVertexBuffer(cbs, _handle, _offset, _size, _stride, alignment);
+                    int stride = (_stride + (alignment - 1)) & -alignment;
+
+                    var buffer = autoBuffer.Get(cbs, _offset, _size).Value;
+
+                    if (gd.Capabilities.SupportsExtendedDynamicState)
+                    {
+                        gd.ExtendedDynamicStateApi.CmdBindVertexBuffers2(
+                            cbs.CommandBuffer,
+                            binding,
+                            1,
+                            buffer,
+                            0,
+                            (ulong)(_size / _stride) * (ulong)stride,
+                            (ulong)stride);
+                    }
+                    else
+                    {
+                        gd.Api.CmdBindVertexBuffers(cbs.CommandBuffer, binding, 1, buffer, 0);
+                    }
+
+                    _buffer = autoBuffer;
+                    state.Internal.VertexBindingDescriptions[DescriptorIndex].Stride = (uint)stride;
+
+                    return;
+                }
+                else
+                {
+                    autoBuffer = gd.BufferManager.GetBuffer(cbs.CommandBuffer, _handle, false, out int _);
+
+                    // The original stride must be reapplied in case it was rewritten.
+                    state.Internal.VertexBindingDescriptions[DescriptorIndex].Stride = (uint)_stride;
+                }
+            }
+
+            if (autoBuffer != null)
+            {
+                var buffer = autoBuffer.Get(cbs, _offset, _size).Value;
+
+                if (gd.Capabilities.SupportsExtendedDynamicState)
+                {
+                    gd.ExtendedDynamicStateApi.CmdBindVertexBuffers2(
+                        cbs.CommandBuffer,
+                        binding,
+                        1,
+                        buffer,
+                        (ulong)_offset,
+                        (ulong)_size,
+                        (ulong)_stride);
+                }
+                else
+                {
+                    gd.Api.CmdBindVertexBuffers(cbs.CommandBuffer, binding, 1, buffer, (ulong)_offset);
+                }
+            }
+        }
+
+        public bool BoundEquals(Auto<DisposableBuffer> buffer)
+        {
+            return _buffer == buffer;
+        }
+
+        public void Dispose()
+        {
+            // Only dispose if this buffer is not refetched on each bind.
+
+            if (_handle == BufferHandle.Null)
+            {
+                _buffer?.DecrementReferenceCount();
+            }
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/VulkanInitialization.cs

@@ -28,6 +28,7 @@ namespace Ryujinx.Graphics.Vulkan
             "VK_EXT_fragment_shader_interlock",
             "VK_EXT_index_type_uint8",
             "VK_EXT_robustness2",
+            "VK_KHR_shader_float16_int8",
             "VK_EXT_shader_subgroup_ballot",
             "VK_EXT_subgroup_size_control",
             "VK_NV_geometry_shader_passthrough"

Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -188,11 +188,22 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PhysicalDeviceRobustness2FeaturesExt
             };
 
+            PhysicalDeviceShaderFloat16Int8FeaturesKHR featuresShaderInt8 = new PhysicalDeviceShaderFloat16Int8FeaturesKHR()
+            {
+                SType = StructureType.PhysicalDeviceShaderFloat16Int8Features
+            };
+
             if (supportedExtensions.Contains("VK_EXT_robustness2"))
             {
                 features2.PNext = &featuresRobustness2;
             }
 
+            if (supportedExtensions.Contains("VK_KHR_shader_float16_int8"))
+            {
+                featuresShaderInt8.PNext = features2.PNext;
+                features2.PNext = &featuresShaderInt8;
+            }
+
             Api.GetPhysicalDeviceFeatures2(_physicalDevice, &features2);
 
             Capabilities = new HardwareCapabilities(
@@ -202,6 +213,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportedExtensions.Contains("VK_EXT_fragment_shader_interlock"),
                 supportedExtensions.Contains("VK_NV_geometry_shader_passthrough"),
                 supportedExtensions.Contains("VK_EXT_subgroup_size_control"),
+                featuresShaderInt8.ShaderInt8,
                 supportedExtensions.Contains(ExtConditionalRendering.ExtensionName),
                 supportedExtensions.Contains(ExtExtendedDynamicState.ExtensionName),
                 features2.Features.MultiViewport,
@@ -506,6 +518,24 @@ namespace Ryujinx.Graphics.Vulkan
             PrintGpuInformation();
         }
 
+        public bool NeedsVertexBufferAlignment(int attrScalarAlignment, out int alignment)
+        {
+            if (Vendor != Vendor.Nvidia)
+            {
+                // Vulkan requires that vertex attributes are globally aligned by their component size,
+                // so buffer strides that don't divide by the largest scalar element are invalid.
+                // Guest applications do this, NVIDIA GPUs are OK with it, others are not.
+
+                alignment = attrScalarAlignment;
+
+                return true;
+            }
+
+            alignment = 1;
+
+            return false;
+        }
+
         public void PreFrame()
         {
             _syncManager.Cleanup();

Ryujinx.HLE/HOS/Services/Audio/AudioRendererManager.cs

@@ -31,11 +31,26 @@ namespace Ryujinx.HLE.HOS.Services.Audio
             return AudioRendererManagerImpl.GetWorkBufferSize(ref parameter);
         }
 
-        public ResultCode OpenAudioRenderer(ServiceCtx context, out IAudioRenderer obj, ref AudioRendererConfiguration parameter, ulong workBufferSize, ulong appletResourceUserId, KTransferMemory workBufferTransferMemory, uint processHandle)
+        public ResultCode OpenAudioRenderer(
+            ServiceCtx context,
+            out IAudioRenderer obj,
+            ref AudioRendererConfiguration parameter,
+            ulong workBufferSize,
+            ulong appletResourceUserId,
+            KTransferMemory workBufferTransferMemory,
+            uint processHandle)
         {
             var memoryManager = context.Process.HandleTable.GetKProcess((int)processHandle).CpuMemory;
 
-            ResultCode result = (ResultCode)_impl.OpenAudioRenderer(out AudioRenderSystem renderer, memoryManager, ref parameter, appletResourceUserId, workBufferTransferMemory.Address, workBufferTransferMemory.Size, processHandle, context.Device.Configuration.AudioVolume);
+            ResultCode result = (ResultCode)_impl.OpenAudioRenderer(
+                out AudioRenderSystem renderer,
+                memoryManager,
+                ref parameter,
+                appletResourceUserId,
+                workBufferTransferMemory.Address,
+                workBufferTransferMemory.Size,
+                processHandle,
+                context.Device.Configuration.AudioVolume);
 
             if (result == ResultCode.Success)
             {

Ryujinx.HLE/HOS/Services/Audio/AudioRendererManagerServer.cs

@@ -34,7 +34,14 @@ namespace Ryujinx.HLE.HOS.Services.Audio
             KTransferMemory workBufferTransferMemory = context.Process.HandleTable.GetObject<KTransferMemory>(transferMemoryHandle);
             uint processHandle = (uint)context.Request.HandleDesc.ToCopy[1];
 
-            ResultCode result = _impl.OpenAudioRenderer(context, out IAudioRenderer renderer, ref parameter, workBufferSize, appletResourceUserId, workBufferTransferMemory, processHandle);
+            ResultCode result = _impl.OpenAudioRenderer(
+                context,
+                out IAudioRenderer renderer,
+                ref parameter,
+                workBufferSize,
+                appletResourceUserId,
+                workBufferTransferMemory,
+                processHandle);
 
             if (result == ResultCode.Success)
             {

Ryujinx.HLE/HOS/Services/Sockets/Bsd/IClient.cs

@@ -336,6 +336,12 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
                 context.Memory.Write(outputBufferPosition + (ulong)(i * Unsafe.SizeOf<PollEventData>()), events[i].Data);
             }
 
+            // In case of non blocking call timeout should not be returned.
+            if (timeout == 0 && errno == LinuxError.ETIMEDOUT)
+            {
+                errno = LinuxError.SUCCESS;
+            }
+
             return WriteBsdResult(context, updateCount, errno);
         }
 
@@ -567,8 +573,11 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
 
             LinuxError errno  = LinuxError.EBADF;
             ISocket    socket = _context.RetrieveSocket(socketFd);
-
             if (socket != null)
+            {
+                errno = LinuxError.ENOTCONN;
+
+                if (socket.RemoteEndPoint != null)
                 {
                     errno = LinuxError.SUCCESS;
 
@@ -576,6 +585,7 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
                     WriteBsdResult(context, 0, errno);
                     context.ResponseData.Write(Unsafe.SizeOf<BsdSockAddr>());
                 }
+            }
 
             return WriteBsdResult(context, 0, errno);
         }
@@ -876,6 +886,91 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
             return WriteBsdResult(context, newSockFd, errno);
         }
 
+
+        [CommandHipc(29)] // 7.0.0+
+        // RecvMMsg(u32 fd, u32 vlen, u32 flags, u32 reserved, nn::socket::TimeVal timeout) -> (i32 ret, u32 bsd_errno, buffer<bytes, 6> message);
+        public ResultCode RecvMMsg(ServiceCtx context)
+        {
+            int            socketFd    = context.RequestData.ReadInt32();
+            int            vlen        = context.RequestData.ReadInt32();
+            BsdSocketFlags socketFlags = (BsdSocketFlags)context.RequestData.ReadInt32();
+            uint           reserved    = context.RequestData.ReadUInt32();
+            TimeVal        timeout     = context.RequestData.ReadStruct<TimeVal>();
+
+            ulong receivePosition = context.Request.ReceiveBuff[0].Position;
+            ulong receiveLength = context.Request.ReceiveBuff[0].Size;
+
+            WritableRegion receiveRegion = context.Memory.GetWritableRegion(receivePosition, (int)receiveLength);
+
+            LinuxError errno  = LinuxError.EBADF;
+            ISocket    socket = _context.RetrieveSocket(socketFd);
+            int        result = -1;
+
+            if (socket != null)
+            {
+                errno = BsdMMsgHdr.Deserialize(out BsdMMsgHdr message, receiveRegion.Memory.Span, vlen);
+
+                if (errno == LinuxError.SUCCESS)
+                {
+                    errno = socket.RecvMMsg(out result, message, socketFlags, timeout);
+
+                    if (errno == LinuxError.SUCCESS)
+                    {
+                        errno = BsdMMsgHdr.Serialize(receiveRegion.Memory.Span, message);
+                    }
+                }
+            }
+
+            if (errno == LinuxError.SUCCESS)
+            {
+                SetResultErrno(socket, result);
+                receiveRegion.Dispose();
+            }
+
+            return WriteBsdResult(context, result, errno);
+        }
+
+        [CommandHipc(30)] // 7.0.0+
+        // SendMMsg(u32 fd, u32 vlen, u32 flags) -> (i32 ret, u32 bsd_errno, buffer<bytes, 6> message);
+        public ResultCode SendMMsg(ServiceCtx context)
+        {
+            int            socketFd    = context.RequestData.ReadInt32();
+            int            vlen        = context.RequestData.ReadInt32();
+            BsdSocketFlags socketFlags = (BsdSocketFlags)context.RequestData.ReadInt32();
+
+            ulong receivePosition = context.Request.ReceiveBuff[0].Position;
+            ulong receiveLength = context.Request.ReceiveBuff[0].Size;
+
+            WritableRegion receiveRegion = context.Memory.GetWritableRegion(receivePosition, (int)receiveLength);
+
+            LinuxError errno  = LinuxError.EBADF;
+            ISocket    socket = _context.RetrieveSocket(socketFd);
+            int        result = -1;
+
+            if (socket != null)
+            {
+                errno = BsdMMsgHdr.Deserialize(out BsdMMsgHdr message, receiveRegion.Memory.Span, vlen);
+
+                if (errno == LinuxError.SUCCESS)
+                {
+                    errno = socket.SendMMsg(out result, message, socketFlags);
+
+                    if (errno == LinuxError.SUCCESS)
+                    {
+                        errno = BsdMMsgHdr.Serialize(receiveRegion.Memory.Span, message);
+                    }
+                }
+            }
+
+            if (errno == LinuxError.SUCCESS)
+            {
+                SetResultErrno(socket, result);
+                receiveRegion.Dispose();
+            }
+
+            return WriteBsdResult(context, result, errno);
+        }
+
         [CommandHipc(31)] // 7.0.0+
         // EventFd(u64 initval, nn::socket::EventFdFlags flags) -> (i32 ret, u32 bsd_errno)
         public ResultCode EventFd(ServiceCtx context)

Ryujinx.HLE/HOS/Services/Sockets/Bsd/ISocket.cs

@@ -25,7 +25,12 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
 
         LinuxError SendTo(out int sendSize, ReadOnlySpan<byte> buffer, int size, BsdSocketFlags flags, IPEndPoint remoteEndPoint);
 
+        LinuxError RecvMMsg(out int vlen, BsdMMsgHdr message, BsdSocketFlags flags, TimeVal timeout);
+
+        LinuxError SendMMsg(out int vlen, BsdMMsgHdr message, BsdSocketFlags flags);
+
         LinuxError GetSocketOption(BsdSocketOption option, SocketOptionLevel level, Span<byte> optionValue);
+
         LinuxError SetSocketOption(BsdSocketOption option, SocketOptionLevel level, ReadOnlySpan<byte> optionValue);
 
         bool Poll(int microSeconds, SelectMode mode);

Ryujinx.HLE/HOS/Services/Sockets/Bsd/Impl/ManagedSocket.cs

@@ -1,5 +1,7 @@
 using Ryujinx.Common.Logging;
 using System;
+using System.Collections.Generic;
+using System.Diagnostics;
 using System.Net;
 using System.Net.Sockets;
 using System.Runtime.InteropServices;
@@ -356,5 +358,165 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
         {
             return Send(out writeSize, buffer, BsdSocketFlags.None);
         }
+
+        private bool CanSupportMMsgHdr(BsdMMsgHdr message)
+        {
+            for (int i = 0; i < message.Messages.Length; i++)
+            {
+                if (message.Messages[i].Name != null ||
+                    message.Messages[i].Control != null)
+                {
+                    return false;
+                }
+            }
+
+            return true;
+        }
+
+        private static IList<ArraySegment<byte>> ConvertMessagesToBuffer(BsdMMsgHdr message)
+        {
+            int segmentCount = 0;
+            int index = 0;
+
+            foreach (BsdMsgHdr msgHeader in message.Messages)
+            {
+                segmentCount += msgHeader.Iov.Length;
+            }
+
+            ArraySegment<byte>[] buffers = new ArraySegment<byte>[segmentCount];
+
+            foreach (BsdMsgHdr msgHeader in message.Messages)
+            {
+                foreach (byte[] iov in msgHeader.Iov)
+                {
+                    buffers[index++] = new ArraySegment<byte>(iov);
+                }
+
+                // Clear the length
+                msgHeader.Length = 0;
+            }
+
+            return buffers;
+        }
+
+        private static void UpdateMessages(out int vlen, BsdMMsgHdr message, int transferedSize)
+        {
+            int bytesLeft = transferedSize;
+            int index = 0;
+
+            while (bytesLeft > 0)
+            {
+                // First ensure we haven't finished all buffers
+                if (index >= message.Messages.Length)
+                {
+                    break;
+                }
+
+                BsdMsgHdr msgHeader = message.Messages[index];
+
+                int possiblyTransferedBytes = 0;
+
+                foreach (byte[] iov in msgHeader.Iov)
+                {
+                    possiblyTransferedBytes += iov.Length;
+                }
+
+                int storedBytes;
+
+                if (bytesLeft > possiblyTransferedBytes)
+                {
+                    storedBytes = possiblyTransferedBytes;
+                    index++;
+                }
+                else
+                {
+                    storedBytes = bytesLeft;
+                }
+
+                msgHeader.Length = (uint)storedBytes;
+                bytesLeft -= storedBytes;
+            }
+
+            Debug.Assert(bytesLeft == 0);
+
+            vlen = index + 1;
+        }
+
+        // TODO: Find a way to support passing the timeout somehow without changing the socket ReceiveTimeout.
+        public LinuxError RecvMMsg(out int vlen, BsdMMsgHdr message, BsdSocketFlags flags, TimeVal timeout)
+        {
+            vlen = 0;
+
+            if (message.Messages.Length == 0)
+            {
+                return LinuxError.SUCCESS;
+            }
+
+            if (!CanSupportMMsgHdr(message))
+            {
+                Logger.Warning?.Print(LogClass.ServiceBsd, $"Unsupported BsdMMsgHdr");
+
+                return LinuxError.EOPNOTSUPP;
+            }
+
+            if (message.Messages.Length == 0)
+            {
+                return LinuxError.SUCCESS;
+            }
+
+            try
+            {
+                int receiveSize = Socket.Receive(ConvertMessagesToBuffer(message), ConvertBsdSocketFlags(flags), out SocketError socketError);
+
+                if (receiveSize > 0)
+                {
+                    UpdateMessages(out vlen, message, receiveSize);
+                }
+
+                return WinSockHelper.ConvertError((WsaError)socketError);
+            }
+            catch (SocketException exception)
+            {
+                return WinSockHelper.ConvertError((WsaError)exception.ErrorCode);
+            }
+        }
+
+        public LinuxError SendMMsg(out int vlen, BsdMMsgHdr message, BsdSocketFlags flags)
+        {
+            vlen = 0;
+
+            if (message.Messages.Length == 0)
+            {
+                return LinuxError.SUCCESS;
+            }
+
+            if (!CanSupportMMsgHdr(message))
+            {
+                Logger.Warning?.Print(LogClass.ServiceBsd, $"Unsupported BsdMMsgHdr");
+
+                return LinuxError.EOPNOTSUPP;
+            }
+
+            if (message.Messages.Length == 0)
+            {
+                return LinuxError.SUCCESS;
+            }
+
+            try
+            {
+                int sendSize = Socket.Send(ConvertMessagesToBuffer(message), ConvertBsdSocketFlags(flags), out SocketError socketError);
+
+                if (sendSize > 0)
+                {
+                    UpdateMessages(out vlen, message, sendSize);
+                }
+
+                return WinSockHelper.ConvertError((WsaError)socketError);
+            }
+            catch (SocketException exception)
+            {
+                return WinSockHelper.ConvertError((WsaError)exception.ErrorCode);
+            }
+        }
     }
 }

Ryujinx.HLE/HOS/Services/Sockets/Bsd/Impl/ManagedSocketPollManager.cs

@@ -38,12 +38,13 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
             {
                 ManagedSocket socket = (ManagedSocket)evnt.FileDescriptor;
 
-                bool isValidEvent = false;
+                bool isValidEvent = evnt.Data.InputEvents == 0;
+
+                errorEvents.Add(socket.Socket);
 
                 if ((evnt.Data.InputEvents & PollEventTypeMask.Input) != 0)
                 {
                     readEvents.Add(socket.Socket);
-                    errorEvents.Add(socket.Socket);
 
                     isValidEvent = true;
                 }
@@ -51,7 +52,6 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
                 if ((evnt.Data.InputEvents & PollEventTypeMask.UrgentInput) != 0)
                 {
                     readEvents.Add(socket.Socket);
-                    errorEvents.Add(socket.Socket);
 
                     isValidEvent = true;
                 }
@@ -59,14 +59,6 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
                 if ((evnt.Data.InputEvents & PollEventTypeMask.Output) != 0)
                 {
                     writeEvents.Add(socket.Socket);
-                    errorEvents.Add(socket.Socket);
-
-                    isValidEvent = true;
-                }
-
-                if ((evnt.Data.InputEvents & PollEventTypeMask.Error) != 0)
-                {
-                    errorEvents.Add(socket.Socket);
 
                     isValidEvent = true;
                 }
@@ -93,7 +85,7 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
             {
                 Socket socket = ((ManagedSocket)evnt.FileDescriptor).Socket;
 
-                PollEventTypeMask outputEvents = 0;
+                PollEventTypeMask outputEvents = evnt.Data.OutputEvents & ~evnt.Data.InputEvents;
 
                 if (errorEvents.Contains(socket))
                 {

Ryujinx.HLE/HOS/Services/Sockets/Bsd/Types/BsdMMsgHdr.cs

@@ -0,0 +1,56 @@
+using System;
+
+namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
+{
+    class BsdMMsgHdr
+    {
+        public BsdMsgHdr[] Messages { get; }
+
+        private BsdMMsgHdr(BsdMsgHdr[] messages)
+        {
+            Messages = messages;
+        }
+
+        public static LinuxError Serialize(Span<byte> rawData, BsdMMsgHdr message)
+        {
+            rawData[0] = 0x8;
+            rawData = rawData[1..];
+
+            for (int index = 0; index < message.Messages.Length; index++)
+            {
+                LinuxError res = BsdMsgHdr.Serialize(ref rawData, message.Messages[index]);
+
+                if (res != LinuxError.SUCCESS)
+                {
+                    return res;
+                }
+            }
+
+            return LinuxError.SUCCESS;
+        }
+
+        public static LinuxError Deserialize(out BsdMMsgHdr message, ReadOnlySpan<byte> rawData, int vlen)
+        {
+            message = null;
+
+            BsdMsgHdr[] messages = new BsdMsgHdr[vlen];
+
+            // Skip "header" byte (Nintendo also ignore it)
+            rawData = rawData[1..];
+
+            for (int index = 0; index < messages.Length; index++)
+            {
+                LinuxError res = BsdMsgHdr.Deserialize(out messages[index], ref rawData);
+
+                if (res != LinuxError.SUCCESS)
+                {
+                    return res;
+                }
+            }
+
+            message = new BsdMMsgHdr(messages);
+
+            return LinuxError.SUCCESS;
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Sockets/Bsd/Types/BsdMsgHdr.cs

@@ -0,0 +1,212 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
+{
+    class BsdMsgHdr
+    {
+        public byte[] Name { get; }
+        public byte[][] Iov { get; }
+        public byte[] Control { get; }
+        public BsdSocketFlags Flags { get; }
+        public uint Length;
+
+        private BsdMsgHdr(byte[] name, byte[][] iov, byte[] control, BsdSocketFlags flags, uint length)
+        {
+            Name = name;
+            Iov = iov;
+            Control = control;
+            Flags = flags;
+            Length = length;
+        }
+
+        public static LinuxError Serialize(ref Span<byte> rawData, BsdMsgHdr message)
+        {
+            int msgNameLength = message.Name == null ? 0 : message.Name.Length;
+            int iovCount = message.Iov == null ? 0 : message.Iov.Length;
+            int controlLength = message.Control == null ? 0 : message.Control.Length;
+            BsdSocketFlags flags = message.Flags;
+
+            if (!MemoryMarshal.TryWrite(rawData, ref msgNameLength))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (msgNameLength > 0)
+            {
+                if (rawData.Length < msgNameLength)
+                {
+                    return LinuxError.EFAULT;
+                }
+
+                message.Name.CopyTo(rawData);
+                rawData = rawData[msgNameLength..];
+            }
+
+            if (!MemoryMarshal.TryWrite(rawData, ref iovCount))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (iovCount > 0)
+            {
+                for (int index = 0; index < iovCount; index++)
+                {
+                    ulong iovLength = (ulong)message.Iov[index].Length;
+
+                    if (!MemoryMarshal.TryWrite(rawData, ref iovLength))
+                    {
+                        return LinuxError.EFAULT;
+                    }
+
+                    rawData = rawData[sizeof(ulong)..];
+
+                    if (iovLength > 0)
+                    {
+                        if ((ulong)rawData.Length < iovLength)
+                        {
+                            return LinuxError.EFAULT;
+                        }
+
+                        message.Iov[index].CopyTo(rawData);
+                        rawData = rawData[(int)iovLength..];
+                    }
+                }
+            }
+
+            if (!MemoryMarshal.TryWrite(rawData, ref controlLength))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (controlLength > 0)
+            {
+                if (rawData.Length < controlLength)
+                {
+                    return LinuxError.EFAULT;
+                }
+
+                message.Control.CopyTo(rawData);
+                rawData = rawData[controlLength..];
+            }
+
+            if (!MemoryMarshal.TryWrite(rawData, ref flags))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(BsdSocketFlags)..];
+
+            if (!MemoryMarshal.TryWrite(rawData, ref message.Length))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            return LinuxError.SUCCESS;
+        }
+
+        public static LinuxError Deserialize(out BsdMsgHdr message, ref ReadOnlySpan<byte> rawData)
+        {
+            byte[] name = null;
+            byte[][] iov = null;
+            byte[] control = null;
+
+            message = null;
+
+            if (!MemoryMarshal.TryRead(rawData, out uint msgNameLength))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (msgNameLength > 0)
+            {
+                if (rawData.Length < msgNameLength)
+                {
+                    return LinuxError.EFAULT;
+                }
+
+                name = rawData[..(int)msgNameLength].ToArray();
+                rawData = rawData[(int)msgNameLength..];
+            }
+
+            if (!MemoryMarshal.TryRead(rawData, out uint iovCount))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (iovCount > 0)
+            {
+                iov = new byte[iovCount][];
+
+                for (int index = 0; index < iov.Length; index++)
+                {
+                    if (!MemoryMarshal.TryRead(rawData, out ulong iovLength))
+                    {
+                        return LinuxError.EFAULT;
+                    }
+
+                    rawData = rawData[sizeof(ulong)..];
+
+                    if (iovLength > 0)
+                    {
+                        if ((ulong)rawData.Length < iovLength)
+                        {
+                            return LinuxError.EFAULT;
+                        }
+
+                        iov[index] = rawData[..(int)iovLength].ToArray();
+                        rawData = rawData[(int)iovLength..];
+                    }
+                }
+            }
+
+            if (!MemoryMarshal.TryRead(rawData, out uint controlLength))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            if (controlLength > 0)
+            {
+                if (rawData.Length < controlLength)
+                {
+                    return LinuxError.EFAULT;
+                }
+
+                control = rawData[..(int)controlLength].ToArray();
+                rawData = rawData[(int)controlLength..];
+            }
+
+            if (!MemoryMarshal.TryRead(rawData, out BsdSocketFlags flags))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(BsdSocketFlags)..];
+
+            if (!MemoryMarshal.TryRead(rawData, out uint length))
+            {
+                return LinuxError.EFAULT;
+            }
+
+            rawData = rawData[sizeof(uint)..];
+
+            message = new BsdMsgHdr(name, iov, control, flags, length);
+
+            return LinuxError.SUCCESS;
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Sockets/Bsd/Types/TimeVal.cs

@@ -0,0 +1,8 @@
+namespace Ryujinx.HLE.HOS.Services.Sockets.Bsd
+{
+    public struct TimeVal
+    {
+        public ulong TvSec;
+        public ulong TvUsec;
+    }
+}

Ryujinx.HLE/HOS/Services/Sockets/Sfdnsres/Types/AddrInfo4.cs

@@ -19,7 +19,7 @@ namespace Ryujinx.HLE.HOS.Services.Sockets.Sfdnsres.Types
         {
             Length  = (byte)Unsafe.SizeOf<Array4<byte>>();
             Family  = (byte)AddressFamily.InterNetwork;
-            Port    = port;
+            Port    = IPAddress.HostToNetworkOrder(port);
             Address = new Array4<byte>();
 
             address.TryWriteBytes(Address.AsSpan(), out _);

Ryujinx.HLE/HOS/Services/SurfaceFlinger/SurfaceFlinger.cs

@@ -35,6 +35,7 @@ namespace Ryujinx.HLE.HOS.Services.SurfaceFlinger
         private long _1msTicks;
 
         private int _swapInterval;
+        private int _swapIntervalDelay;
 
         private readonly object Lock = new object();
 
@@ -91,7 +92,7 @@ namespace Ryujinx.HLE.HOS.Services.SurfaceFlinger
             }
             else
             {
-                _ticksPerFrame = Stopwatch.Frequency / (TargetFps / _swapInterval);
+                _ticksPerFrame = Stopwatch.Frequency / TargetFps;
             }
         }
 
@@ -321,9 +322,15 @@ namespace Ryujinx.HLE.HOS.Services.SurfaceFlinger
                     lastTicks = ticks;
 
                     if (_ticks >= _ticksPerFrame)
+                    {
+                        if (_swapIntervalDelay-- == 0)
                         {
                             Compose();
 
+                            // When a frame is presented, delay the next one by its swap interval value.
+                            _swapIntervalDelay = Math.Max(0, _swapInterval - 1);
+                        }
+
                         _device.System?.SignalVsync();
 
                         // Apply a maximum bound of 3 frames to the tick remainder, in case some event causes Ryujinx to pause for a long time or messes with the timer.

Ryujinx.Tests/Cpu/CpuTestSimd32.cs

@@ -12,12 +12,22 @@ namespace Ryujinx.Tests.Cpu
 #if Simd32
 
 #region "ValueSource (Opcodes)"
-        private static uint[] _Vabs_Vneg_V_()
+        private static uint[] _Vabs_Vneg_Vpaddl_I_()
         {
             return new uint[]
             {
                 0xf3b10300u, // VABS.S8   D0, D0
-                0xf3b10380u  // VNEG.S8 D0, D0
+                0xf3b10380u, // VNEG.S8   D0, D0
+                0xf3b00200u  // VPADDL.S8 D0, D0
+            };
+        }
+
+        private static uint[] _Vabs_Vneg_F_()
+        {
+            return new uint[]
+            {
+                0xf3b90700u, // VABS.F32 D0, D0
+                0xf3b90780u  // VNEG.F32 D0, D0
             };
         }
 #endregion
@@ -201,7 +211,7 @@ namespace Ryujinx.Tests.Cpu
         }
 
         [Test, Pairwise]
-        public void Vabs_Vneg_V_S8_S16_S32([ValueSource("_Vabs_Vneg_V_")] uint opcode,
+        public void Vabs_Vneg_Vpaddl_V_I([ValueSource("_Vabs_Vneg_Vpaddl_I_")] uint opcode,
                                          [Range(0u, 3u)] uint rd,
                                          [Range(0u, 3u)] uint rm,
                                          [ValueSource("_8B4H2S_")] [Random(RndCnt)] ulong z,
@@ -209,32 +219,6 @@ namespace Ryujinx.Tests.Cpu
                                          [Values(0u, 1u, 2u)] uint size, // <S8, S16, S32>
                                          [Values] bool q)
         {
-            const bool f = false;
-
-            Vabs_Vneg_V(opcode, rd, rm, z, b, size, f, q);
-        }
-
-        [Test, Pairwise]
-        public void Vabs_Vneg_V_F32([ValueSource("_Vabs_Vneg_V_")] uint opcode,
-                                    [Range(0u, 3u)] uint rd,
-                                    [Range(0u, 3u)] uint rm,
-                                    [ValueSource("_2S_F_")] ulong z,
-                                    [ValueSource("_2S_F_")] ulong b,
-                                    [Values] bool q)
-        {
-            const uint size = 0b10; // <F32>
-            const bool f = true;
-
-            Vabs_Vneg_V(opcode, rd, rm, z, b, size, f, q);
-        }
-
-        private void Vabs_Vneg_V(uint opcode, uint rd, uint rm, ulong z, ulong b, uint size, bool f, bool q)
-        {
-            if (f)
-            {
-                opcode |= 1 << 10;
-            }
-
             if (q)
             {
                 opcode |= 1 << 6;
@@ -256,6 +240,33 @@ namespace Ryujinx.Tests.Cpu
             CompareAgainstUnicorn();
         }
 
+        [Test, Pairwise]
+        public void Vabs_Vneg_V_F32([ValueSource("_Vabs_Vneg_F_")] uint opcode,
+                                    [Range(0u, 3u)] uint rd,
+                                    [Range(0u, 3u)] uint rm,
+                                    [ValueSource("_2S_F_")] ulong z,
+                                    [ValueSource("_2S_F_")] ulong b,
+                                    [Values] bool q)
+        {
+            if (q)
+            {
+                opcode |= 1 << 6;
+
+                rd >>= 1; rd <<= 1;
+                rm >>= 1; rm <<= 1;
+            }
+
+            opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
+
+            V128 v0 = MakeVectorE0E1(z, ~z);
+            V128 v1 = MakeVectorE0E1(b, ~b);
+
+            SingleOpcode(opcode, v0: v0, v1: v1);
+
+            CompareAgainstUnicorn();
+        }
+
         [Test, Pairwise, Description("VCNT.8 D0, D0 | VCNT.8 Q0, Q0")]
         public void Vcnt([Values(0u, 1u)] uint rd,
                          [Values(0u, 1u)] uint rm,
@@ -283,6 +294,32 @@ namespace Ryujinx.Tests.Cpu
 
             CompareAgainstUnicorn();
         }
+
+        [Test, Pairwise]
+        public void Vmovn_V([Range(0u, 3u)] uint rd,
+                            [Range(0u, 3u)] uint rm,
+                            [ValueSource("_8B4H2S_")] [Random(RndCnt)] ulong z,
+                            [ValueSource("_8B4H2S_")] [Random(RndCnt)] ulong b,
+                            [Values(0u, 1u, 2u, 3u)] uint op,
+                            [Values(0u, 1u, 2u)] uint size) // <S8, S16, S32>
+        {
+            rm >>= 1; rm <<= 1;
+
+            uint opcode = 0xf3b20200u; // VMOVN.S16 D0, Q0
+
+            opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
+
+            opcode |= (op & 0x3) << 6;
+            opcode |= (size & 0x3) << 18;
+
+            V128 v0 = MakeVectorE0E1(z, ~z);
+            V128 v1 = MakeVectorE0E1(b, ~b);
+
+            SingleOpcode(opcode, v0: v0, v1: v1);
+
+            CompareAgainstUnicorn();
+        }
 #endif
     }
 }

Ryujinx.Tests/Cpu/CpuTestSimdReg32.cs

@@ -12,11 +12,13 @@ namespace Ryujinx.Tests.Cpu
 #if SimdReg32
 
 #region "ValueSource (Opcodes)"
-        private static uint[] _V_Add_Sub_Wide_I_()
+        private static uint[] _V_Add_Sub_Long_Wide_I_()
         {
             return new uint[]
             {
+                0xf2800000u, // VADDL.S8 Q0, D0, D0
                 0xf2800100u, // VADDW.S8 Q0, Q0, D0
+                0xf2800200u, // VSUBL.S8 Q0, D0, D0
                 0xf2800300u  // VSUBW.S8 Q0, Q0, D0
             };
         }
@@ -74,6 +76,15 @@ namespace Ryujinx.Tests.Cpu
             };
         }
 
+        private static uint[] _Vmlal_Vmlsl_V_I_()
+        {
+            return new uint[]
+            {
+                0xf2800800u, // VMLAL.S8 Q0, D0, D0
+                0xf2800a00u  // VMLSL.S8 Q0, D0, D0
+            };
+        }
+
         private static uint[] _Vp_Add_Max_Min_F_()
         {
             return new uint[]
@@ -84,16 +95,30 @@ namespace Ryujinx.Tests.Cpu
             };
         }
 
-        // VPADD does not have an unsigned flag, so we check the opcode before setting it.
-        private static uint VpaddI8 = 0xf2000b10u; // VPADD.I8 D0, D0, D0
-
-        private static uint[] _Vp_Add_Max_Min_I_()
+        private static uint[] _Vp_Add_I_()
+        {
+            return new uint[]
+            {
+                0xf2000b10u // VPADD.I8 D0, D0, D0
+            };
+        }
+
+        private static uint[] _V_Pmax_Pmin_Rhadd_I_()
         {
             return new uint[]
             {
-                VpaddI8,
                 0xf2000a00u, // VPMAX .S8 D0, D0, D0
-                0xf2000a10u  // VPMIN.S8 D0, D0, D0
+                0xf2000a10u, // VPMIN .S8 D0, D0, D0
+                0xf2000100u, // VRHADD.S8 D0, D0, D0
+            };
+        }
+
+        private static uint[] _Vq_Add_Sub_I_()
+        {
+            return new uint[]
+            {
+                0xf2000050u, // VQADD.S8 Q0, Q0, Q0
+                0xf2000250u  // VQSUB.S8 Q0, Q0, Q0
             };
         }
 #endregion
@@ -350,7 +375,7 @@ namespace Ryujinx.Tests.Cpu
 
         [Explicit]
         [Test, Pairwise, Description("VADD.f32 V0, V0, V0")]
-        public void Vadd_f32([Values(0u)] uint rd,
+        public void Vadd_F32([Values(0u)] uint rd,
                              [Values(0u, 1u)] uint rn,
                              [Values(0u, 2u)] uint rm,
                              [ValueSource("_2S_F_")] ulong z0,
@@ -384,7 +409,7 @@ namespace Ryujinx.Tests.Cpu
         }
 
         [Test, Pairwise]
-        public void V_Add_Sub_Wide_I([ValueSource("_V_Add_Sub_Wide_I_")] uint opcode,
+        public void V_Add_Sub_Long_Wide_I([ValueSource("_V_Add_Sub_Long_Wide_I_")] uint opcode,
                                           [Range(0u, 5u)] uint rd,
                                           [Range(0u, 5u)] uint rn,
                                           [Range(0u, 5u)] uint rm,
@@ -566,8 +591,9 @@ namespace Ryujinx.Tests.Cpu
             CompareAgainstUnicorn();
         }
 
-        [Test, Pairwise, Description("VMLSL.<type><size> <Vd>, <Vn>, <Vm>")]
-        public void Vmlsl_I([Values(0u)] uint rd,
+        [Test, Pairwise]
+        public void Vmlal_Vmlsl_I([ValueSource(nameof(_Vmlal_Vmlsl_V_I_))] uint opcode,
+                                  [Values(0u)] uint rd,
                                   [Values(1u, 0u)] uint rn,
                                   [Values(2u, 0u)] uint rm,
                                   [Values(0u, 1u, 2u)] uint size,
@@ -576,8 +602,6 @@ namespace Ryujinx.Tests.Cpu
                                   [Random(RndCnt)] ulong b,
                                   [Values] bool u)
         {
-            uint opcode = 0xf2800a00u; // VMLSL.S8 Q0, D0, D0
-
             opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
             opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
             opcode |= ((rn & 0xf) << 16) | ((rn & 0x10) << 3);
@@ -736,7 +760,32 @@ namespace Ryujinx.Tests.Cpu
         }
 
         [Test, Pairwise]
-        public void Vp_Add_Max_Min_I([ValueSource("_Vp_Add_Max_Min_I_")] uint opcode,
+        public void Vp_Add_I([ValueSource("_Vp_Add_I_")] uint opcode,
+                             [Values(0u)] uint rd,
+                             [Range(0u, 5u)] uint rn,
+                             [Range(0u, 5u)] uint rm,
+                             [Values(0u, 1u, 2u)] uint size,
+                             [Random(RndCnt)] ulong z,
+                             [Random(RndCnt)] ulong a,
+                             [Random(RndCnt)] ulong b)
+        {
+            opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
+            opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            opcode |= ((rn & 0xf) << 16) | ((rn & 0x10) << 3);
+
+            opcode |= size << 20;
+
+            V128 v0 = MakeVectorE0E1(z, z);
+            V128 v1 = MakeVectorE0E1(a, z);
+            V128 v2 = MakeVectorE0E1(b, z);
+
+            SingleOpcode(opcode, v0: v0, v1: v1, v2: v2);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise]
+        public void V_Pmax_Pmin_Rhadd_I([ValueSource("_V_Pmax_Pmin_Rhadd_I_")] uint opcode,
                                         [Values(0u)] uint rd,
                                         [Range(0u, 5u)] uint rn,
                                         [Range(0u, 5u)] uint rm,
@@ -746,7 +795,7 @@ namespace Ryujinx.Tests.Cpu
                                         [Random(RndCnt)] ulong b,
                                         [Values] bool u)
         {
-            if (u && opcode != VpaddI8)
+            if (u)
             {
                 opcode |= 1 << 24;
             }
@@ -765,6 +814,71 @@ namespace Ryujinx.Tests.Cpu
 
             CompareAgainstUnicorn();
         }
+
+        [Test, Pairwise]
+        public void Vq_Add_Sub_I([ValueSource("_Vq_Add_Sub_I_")] uint opcode,
+                                 [Range(0u, 5u)] uint rd,
+                                 [Range(0u, 5u)] uint rn,
+                                 [Range(0u, 5u)] uint rm,
+                                 [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong z,
+                                 [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong a,
+                                 [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong b,
+                                 [Values(0u, 1u, 2u)] uint size, // <SU8, SU16, SU32>
+                                 [Values] bool u) // <S, U>
+        {
+            if (u)
+            {
+                opcode |= 1 << 24;
+            }
+
+            rd >>= 1; rd <<= 1;
+            rn >>= 1; rn <<= 1;
+            rm >>= 1; rm <<= 1;
+
+            opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            opcode |= ((rn & 0xf) << 16) | ((rn & 0x10) << 3);
+            opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
+
+            opcode |= (size & 0x3) << 20;
+
+            V128 v0 = MakeVectorE0E1(z, ~z);
+            V128 v1 = MakeVectorE0E1(a, ~a);
+            V128 v2 = MakeVectorE0E1(b, ~b);
+
+            SingleOpcode(opcode, v0: v0, v1: v1, v2: v2);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("VQDMULH.<S16, S32> <Qd>, <Qn>, <Qm>")]
+        public void Vqdmulh_I([Range(0u, 5u)] uint rd,
+                              [Range(0u, 5u)] uint rn,
+                              [Range(0u, 5u)] uint rm,
+                              [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong z,
+                              [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong a,
+                              [ValueSource("_8B4H2S1D_")] [Random(RndCnt)] ulong b,
+                              [Values(1u, 2u)] uint size) // <S16, S32>
+        {
+            rd >>= 1; rd <<= 1;
+            rn >>= 1; rn <<= 1;
+            rm >>= 1; rm <<= 1;
+
+            uint opcode = 0xf2100b40u & ~(3u << 20); // VQDMULH.S16 Q0, Q0, Q0
+
+            opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            opcode |= ((rn & 0xf) << 16) | ((rn & 0x10) << 3);
+            opcode |= ((rm & 0xf) << 0)  | ((rm & 0x10) << 1);
+
+            opcode |= (size & 0x3) << 20;
+
+            V128 v0 = MakeVectorE0E1(z, ~z);
+            V128 v1 = MakeVectorE0E1(a, ~a);
+            V128 v2 = MakeVectorE0E1(b, ~b);
+
+            SingleOpcode(opcode, v0: v0, v1: v1, v2: v2);
+
+            CompareAgainstUnicorn();
+        }
 #endif
     }
 }

Ryujinx.Tests/Cpu/CpuTestSimdShImm32.cs

@@ -41,9 +41,10 @@ namespace Ryujinx.Tests.Cpu
         {
             return new uint[]
             {
+                0xf2880010u, // VSHR.S8  D0, D0, #8
                 0xf2880110u, // VSRA.S8  D0, D0, #8
                 0xf2880210u, // VRSHR.S8 D0, D0, #8
-                0xf2880010u  // VSHR.S8  D0, D0, #8
+                0xf2880310u  // VRSRA.S8 D0, D0, #8
             };
         }
 
@@ -51,9 +52,10 @@ namespace Ryujinx.Tests.Cpu
         {
             return new uint[]
             {
+                0xf2900010u, // VSHR.S16  D0, D0, #16
                 0xf2900110u, // VSRA.S16  D0, D0, #16
                 0xf2900210u, // VRSHR.S16 D0, D0, #16
-                0xf2900010u  // VSHR.S16  D0, D0, #16
+                0xf2900310u  // VRSRA.S16 D0, D0, #16
             };
         }
 
@@ -61,9 +63,10 @@ namespace Ryujinx.Tests.Cpu
         {
             return new uint[]
             {
+                0xf2a00010u, // VSHR.S32  D0, D0, #32
                 0xf2a00110u, // VSRA.S32  D0, D0, #32
                 0xf2a00210u, // VRSHR.S32 D0, D0, #32
-                0xf2a00010u  // VSHR.S32  D0, D0, #32
+                0xf2a00310u  // VRSRA.S32 D0, D0, #32
             };
         }
 
@@ -76,6 +79,25 @@ namespace Ryujinx.Tests.Cpu
                 0xf2800090u  // VSHR.S64  D0, D0, #64
             };
         }
+
+        private static uint[] _Vqshrn_Vqrshrn_Vrshrn_Imm_()
+        {
+            return new uint[]
+            {
+                0xf2800910u, // VORR.I16 D0, #0 (immediate value changes it into QSHRN)
+                0xf2800950u, // VORR.I16 Q0, #0 (immediate value changes it into QRSHRN)
+                0xf2800850u  // VMOV.I16 Q0, #0 (immediate value changes it into RSHRN)
+            };
+        }
+
+        private static uint[] _Vqshrun_Vqrshrun_Imm_()
+        {
+            return new uint[]
+            {
+                0xf3800810u, // VMOV.I16 D0, #0x80 (immediate value changes it into QSHRUN)
+                0xf3800850u  // VMOV.I16 Q0, #0x80 (immediate value changes it into QRSHRUN)
+            };
+        }
 #endregion
 
         private const int RndCnt = 2;
@@ -230,8 +252,9 @@ namespace Ryujinx.Tests.Cpu
             CompareAgainstUnicorn();
         }
 
-        [Test, Pairwise, Description("VQRSHRN.<type><size> <Vd>, <Vm>, #<imm>")]
-        public void Vqrshrn_Imm([Values(0u, 1u)] uint rd,
+        [Test, Pairwise]
+        public void Vqshrn_Vqrshrn_Vrshrn_Imm([ValueSource("_Vqshrn_Vqrshrn_Vrshrn_Imm_")] uint opcode,
+                                              [Values(0u, 1u)] uint rd,
                                               [Values(2u, 0u)] uint rm,
                                               [Values(0u, 1u, 2u)] uint size,
                                               [Random(RndCntShiftImm)] [Values(0u)] uint shiftImm,
@@ -240,8 +263,6 @@ namespace Ryujinx.Tests.Cpu
                                               [Random(RndCnt)] ulong b,
                                               [Values] bool u)
         {
-            uint opcode = 0xf2800950u; // VORR.I16 Q0, #0 (immediate value changes it into QRSHRN)
-
             uint imm = 1u << ((int)size + 3);
             imm |= shiftImm & (imm - 1);
 
@@ -265,8 +286,9 @@ namespace Ryujinx.Tests.Cpu
             CompareAgainstUnicorn(fpsrMask: Fpsr.Qc);
         }
 
-        [Test, Pairwise, Description("VQRSHRUN.<type><size> <Vd>, <Vm>, #<imm>")]
-        public void Vqrshrun_Imm([Values(0u, 1u)] uint rd,
+        [Test, Pairwise]
+        public void Vqshrun_Vqrshrun_Imm([ValueSource("_Vqshrun_Vqrshrun_Imm_")] uint opcode,
+                                         [Values(0u, 1u)] uint rd,
                                          [Values(2u, 0u)] uint rm,
                                          [Values(0u, 1u, 2u)] uint size,
                                          [Random(RndCntShiftImm)] [Values(0u)] uint shiftImm,
@@ -274,8 +296,6 @@ namespace Ryujinx.Tests.Cpu
                                          [Random(RndCnt)] ulong a,
                                          [Random(RndCnt)] ulong b)
         {
-            uint opcode = 0xf3800850u; // VMOV.I16 Q0, #0x80 (immediate value changes it into QRSHRUN)
-
             uint imm = 1u << ((int)size + 3);
             imm |= shiftImm & (imm - 1);