Add Tbl_V Sse opt. with Tests. (#651)

* Add v4, v5, v30, v31 required for Tbl_V Tests.

* Add Tests for Tbl_V.

* Add Tbl_V Sse opt..

* Nit.

* Small opt. on comparison constant vector.

* Nit.

* Add EmitLd/Stvectmp2/3.

* Nit.

ChocolArm64/Instructions/InstEmitSimdMove.cs

@@ -355,11 +355,69 @@ namespace ChocolArm64.Instructions
         {
             OpCodeSimdTbl64 op = (OpCodeSimdTbl64)context.CurrOp;
 
+            if (Optimizations.UseSsse3)
+            {
+                Type[] typesCmpSflSub = new Type[] { typeof(Vector128<sbyte>), typeof(Vector128<sbyte>) };
+                Type[] typesOr        = new Type[] { typeof(Vector128<long> ), typeof(Vector128<long> ) };
+                Type[] typesSav       = new Type[] { typeof(long) };
+
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
+
+                context.EmitLdc_I8(0x0F0F0F0F0F0F0F0FL);
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                context.EmitStvectmp2();
+                context.EmitLdvectmp2();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), typesCmpSflSub));
+
+                context.EmitLdvec(op.Rm);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
+
+                context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesCmpSflSub));
+
+                for (int index = 1; index < op.Size; index++)
+                {
+                    context.EmitLdvec((op.Rn + index) & 0x1F);
+                    context.EmitLdvec(op.Rm);
+
+                    context.EmitLdc_I8(0x1010101010101010L * index);
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesCmpSflSub));
+
+                    context.EmitStvectmp();
+                    context.EmitLdvectmp();
+
+                    context.EmitLdvectmp2();
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), typesCmpSflSub));
+
+                    context.EmitLdvectmp();
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
+
+                    context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesCmpSflSub));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
+                }
+
+                context.EmitStvec(op.Rd);
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    EmitVectorZeroUpper(context, op.Rd);
+                }
+            }
+            else
+            {
                 context.EmitLdvec(op.Rm);
 
                 for (int index = 0; index < op.Size; index++)
                 {
-                context.EmitLdvec((op.Rn + index) & 0x1f);
+                    context.EmitLdvec((op.Rn + index) & 0x1F);
                 }
 
                 switch (op.Size)
@@ -385,6 +443,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitStvec(op.Rd);
             }
+        }
 
         public static void Trn1_V(ILEmitterCtx context)
         {

ChocolArm64/Translation/ILEmitterCtx.cs

@@ -61,7 +61,9 @@ namespace ChocolArm64.Translation
 
         //Vectors are part of another "set" of locals.
         private const int VecGpTmp1Index   = ReservedLocalsCount + 0;
-        private const int UserVecTempStart = ReservedLocalsCount + 1;
+        private const int VecGpTmp2Index   = ReservedLocalsCount + 1;
+        private const int VecGpTmp3Index   = ReservedLocalsCount + 2;
+        private const int UserVecTempStart = ReservedLocalsCount + 3;
 
         private static int _userIntTempCount;
         private static int _userVecTempCount;
@@ -629,6 +631,12 @@ namespace ChocolArm64.Translation
         public void EmitLdvectmp() => EmitLdvec(VecGpTmp1Index);
         public void EmitStvectmp() => EmitStvec(VecGpTmp1Index);
 
+        public void EmitLdvectmp2() => EmitLdvec(VecGpTmp2Index);
+        public void EmitStvectmp2() => EmitStvec(VecGpTmp2Index);
+
+        public void EmitLdvectmp3() => EmitLdvec(VecGpTmp3Index);
+        public void EmitStvectmp3() => EmitStvec(VecGpTmp3Index);
+
         public void EmitLdint(int index) => Ldloc(index, VarType.Int);
         public void EmitStint(int index) => Stloc(index, VarType.Int);
 

Ryujinx.Tests/Cpu/CpuTest.cs

@@ -97,6 +97,10 @@ namespace Ryujinx.Tests.Cpu
                                       Vector128<float> v1  = default(Vector128<float>),
                                       Vector128<float> v2  = default(Vector128<float>),
                                       Vector128<float> v3  = default(Vector128<float>),
+                                      Vector128<float> v4  = default(Vector128<float>),
+                                      Vector128<float> v5  = default(Vector128<float>),
+                                      Vector128<float> v30 = default(Vector128<float>),
+                                      Vector128<float> v31 = default(Vector128<float>),
                                       bool overflow = false, bool carry = false, bool zero = false, bool negative = false,
                                       int fpcr = 0x0, int fpsr = 0x0)
         {
@@ -111,6 +115,10 @@ namespace Ryujinx.Tests.Cpu
             _thread.ThreadState.V1  = v1;
             _thread.ThreadState.V2  = v2;
             _thread.ThreadState.V3  = v3;
+            _thread.ThreadState.V4  = v4;
+            _thread.ThreadState.V5  = v5;
+            _thread.ThreadState.V30 = v30;
+            _thread.ThreadState.V31 = v31;
 
             _thread.ThreadState.Overflow = overflow;
             _thread.ThreadState.Carry    = carry;
@@ -133,6 +141,10 @@ namespace Ryujinx.Tests.Cpu
                 _unicornEmu.Q[1]  = v1;
                 _unicornEmu.Q[2]  = v2;
                 _unicornEmu.Q[3]  = v3;
+                _unicornEmu.Q[4]  = v4;
+                _unicornEmu.Q[5]  = v5;
+                _unicornEmu.Q[30] = v30;
+                _unicornEmu.Q[31] = v31;
 
                 _unicornEmu.OverflowFlag = overflow;
                 _unicornEmu.CarryFlag    = carry;
@@ -169,13 +181,17 @@ namespace Ryujinx.Tests.Cpu
                                               Vector128<float> v1  = default(Vector128<float>),
                                               Vector128<float> v2  = default(Vector128<float>),
                                               Vector128<float> v3  = default(Vector128<float>),
+                                              Vector128<float> v4  = default(Vector128<float>),
+                                              Vector128<float> v5  = default(Vector128<float>),
+                                              Vector128<float> v30 = default(Vector128<float>),
+                                              Vector128<float> v31 = default(Vector128<float>),
                                               bool overflow = false, bool carry = false, bool zero = false, bool negative = false,
                                               int fpcr = 0x0, int fpsr = 0x0)
         {
             Opcode(opcode);
             Opcode(0xD4200000); // BRK #0
             Opcode(0xD65F03C0); // RET
-            SetThreadState(x0, x1, x2, x3, x31, v0, v1, v2, v3, overflow, carry, zero, negative, fpcr, fpsr);
+            SetThreadState(x0, x1, x2, x3, x31, v0, v1, v2, v3, v4, v5, v30, v31, overflow, carry, zero, negative, fpcr, fpsr);
             ExecuteOpcodes();
 
             return GetThreadState();

Ryujinx.Tests/Cpu/CpuTestSimdTbl.cs

@@ -0,0 +1,315 @@
+#define SimdTbl
+
+using NUnit.Framework;
+
+using System.Collections.Generic;
+using System.Runtime.Intrinsics;
+
+namespace Ryujinx.Tests.Cpu
+{
+    [Category("SimdTbl")]
+    public sealed class CpuTestSimdTbl : CpuTest
+    {
+#if SimdTbl
+
+#region "Helper methods"
+        private static ulong GenIdxsForTbls(int regs)
+        {
+            const byte idxInRngMin  = (byte)0;
+                  byte idxInRngMax  = (byte)((16 * regs) - 1);
+                  byte idxOutRngMin = (byte) (16 * regs);
+            const byte idxOutRngMax = (byte)255;
+
+            ulong idxs = 0ul;
+
+            for (int cnt = 1; cnt <= 8; cnt++)
+            {
+                ulong idxInRng  = (ulong)TestContext.CurrentContext.Random.NextByte(idxInRngMin,  idxInRngMax);
+                ulong idxOutRng = (ulong)TestContext.CurrentContext.Random.NextByte(idxOutRngMin, idxOutRngMax);
+
+                ulong idx = TestContext.CurrentContext.Random.NextBool() ? idxInRng : idxOutRng;
+
+                idxs = (idxs << 8) | idx;
+            }
+
+            return idxs;
+        }
+#endregion
+
+#region "ValueSource (Types)"
+        private static ulong[] _8B_()
+        {
+            return new ulong[] { 0x0000000000000000ul, 0x7F7F7F7F7F7F7F7Ful,
+                                 0x8080808080808080ul, 0xFFFFFFFFFFFFFFFFul };
+        }
+
+        private static IEnumerable<ulong> _GenIdxsForTbl1_()
+        {
+            yield return 0x0000000000000000ul;
+            yield return 0x7F7F7F7F7F7F7F7Ful;
+            yield return 0x8080808080808080ul;
+            yield return 0xFFFFFFFFFFFFFFFFul;
+
+            for (int cnt = 1; cnt <= RndCntIdxs; cnt++)
+            {
+                yield return GenIdxsForTbls(regs: 1);
+            }
+        }
+
+        private static IEnumerable<ulong> _GenIdxsForTbl2_()
+        {
+            yield return 0x0000000000000000ul;
+            yield return 0x7F7F7F7F7F7F7F7Ful;
+            yield return 0x8080808080808080ul;
+            yield return 0xFFFFFFFFFFFFFFFFul;
+
+            for (int cnt = 1; cnt <= RndCntIdxs; cnt++)
+            {
+                yield return GenIdxsForTbls(regs: 2);
+            }
+        }
+
+        private static IEnumerable<ulong> _GenIdxsForTbl3_()
+        {
+            yield return 0x0000000000000000ul;
+            yield return 0x7F7F7F7F7F7F7F7Ful;
+            yield return 0x8080808080808080ul;
+            yield return 0xFFFFFFFFFFFFFFFFul;
+
+            for (int cnt = 1; cnt <= RndCntIdxs; cnt++)
+            {
+                yield return GenIdxsForTbls(regs: 3);
+            }
+        }
+
+        private static IEnumerable<ulong> _GenIdxsForTbl4_()
+        {
+            yield return 0x0000000000000000ul;
+            yield return 0x7F7F7F7F7F7F7F7Ful;
+            yield return 0x8080808080808080ul;
+            yield return 0xFFFFFFFFFFFFFFFFul;
+
+            for (int cnt = 1; cnt <= RndCntIdxs; cnt++)
+            {
+                yield return GenIdxsForTbls(regs: 4);
+            }
+        }
+#endregion
+
+#region "ValueSource (Opcodes)"
+        private static uint[] _SingleRegTbl_V_8B_16B_()
+        {
+            return new uint[]
+            {
+                0x0E000000u, // TBL V0.8B, { V0.16B }, V0.8B
+            };
+        }
+
+        private static uint[] _TwoRegTbl_V_8B_16B_()
+        {
+            return new uint[]
+            {
+                0x0E002000u, // TBL V0.8B, { V0.16B, V1.16B }, V0.8B
+            };
+        }
+
+        private static uint[] _ThreeRegTbl_V_8B_16B_()
+        {
+            return new uint[]
+            {
+                0x0E004000u, // TBL V0.8B, { V0.16B, V1.16B, V2.16B }, V0.8B
+            };
+        }
+
+        private static uint[] _FourRegTbl_V_8B_16B_()
+        {
+            return new uint[]
+            {
+                0x0E006000u, // TBL V0.8B, { V0.16B, V1.16B, V2.16B, V3.16B }, V0.8B
+            };
+        }
+#endregion
+
+        private const int RndCntTbls = 2;
+        private const int RndCntIdxs = 2;
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B }, <Vm>.<Ta>")]
+        public void SingleRegTbl_V_8B_16B([ValueSource("_SingleRegTbl_V_8B_16B_")] uint opcodes,
+                                          [Values(0u)] uint rd,
+                                          [Values(1u)] uint rn,
+                                          [Values(2u)] uint rm,
+                                          [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                          [ValueSource("_GenIdxsForTbl1_")] ulong indexes,
+                                          [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v0 = MakeVectorE0E1(z, z);
+            Vector128<float> v1 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v2 = MakeVectorE0E1(indexes, q == 1u ? indexes : 0ul);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B }, <Vm>.<Ta>")]
+        public void TwoRegTbl_V_8B_16B([ValueSource("_TwoRegTbl_V_8B_16B_")] uint opcodes,
+                                       [Values(0u)] uint rd,
+                                       [Values(1u)] uint rn,
+                                       [Values(3u)] uint rm,
+                                       [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                       [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                       [ValueSource("_GenIdxsForTbl2_")] ulong indexes,
+                                       [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v0 = MakeVectorE0E1(z, z);
+            Vector128<float> v1 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v2 = MakeVectorE0E1(table1, table1);
+            Vector128<float> v3 = MakeVectorE0E1(indexes, q == 1u ? indexes : 0ul);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2, v3: v3);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B }, <Vm>.<Ta>")]
+        public void Mod_TwoRegTbl_V_8B_16B([ValueSource("_TwoRegTbl_V_8B_16B_")] uint opcodes,
+                                           [Values(30u, 1u)] uint rd,
+                                           [Values(31u)]     uint rn,
+                                           [Values(1u, 30u)] uint rm,
+                                           [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                           [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                           [ValueSource("_GenIdxsForTbl2_")] ulong indexes,
+                                           [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v30 = MakeVectorE0E1(z, z);
+            Vector128<float> v31 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v0  = MakeVectorE0E1(table1, table1);
+            Vector128<float> v1  = MakeVectorE0E1(indexes, indexes);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v30: v30, v31: v31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B, <Vn+2>.16B }, <Vm>.<Ta>")]
+        public void ThreeRegTbl_V_8B_16B([ValueSource("_ThreeRegTbl_V_8B_16B_")] uint opcodes,
+                                         [Values(0u)] uint rd,
+                                         [Values(1u)] uint rn,
+                                         [Values(4u)] uint rm,
+                                         [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                         [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                         [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table2,
+                                         [ValueSource("_GenIdxsForTbl3_")] ulong indexes,
+                                         [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v0 = MakeVectorE0E1(z, z);
+            Vector128<float> v1 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v2 = MakeVectorE0E1(table1, table1);
+            Vector128<float> v3 = MakeVectorE0E1(table2, table2);
+            Vector128<float> v4 = MakeVectorE0E1(indexes, q == 1u ? indexes : 0ul);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2, v3: v3, v4: v4);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B, <Vn+2>.16B }, <Vm>.<Ta>")]
+        public void Mod_ThreeRegTbl_V_8B_16B([ValueSource("_ThreeRegTbl_V_8B_16B_")] uint opcodes,
+                                             [Values(30u, 2u)] uint rd,
+                                             [Values(31u)]     uint rn,
+                                             [Values(2u, 30u)] uint rm,
+                                             [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                             [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                             [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table2,
+                                             [ValueSource("_GenIdxsForTbl3_")] ulong indexes,
+                                             [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v30 = MakeVectorE0E1(z, z);
+            Vector128<float> v31 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v0  = MakeVectorE0E1(table1, table1);
+            Vector128<float> v1  = MakeVectorE0E1(table2, table2);
+            Vector128<float> v2  = MakeVectorE0E1(indexes, indexes);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2, v30: v30, v31: v31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B, <Vn+2>.16B, <Vn+3>.16B }, <Vm>.<Ta>")]
+        public void FourRegTbl_V_8B_16B([ValueSource("_FourRegTbl_V_8B_16B_")] uint opcodes,
+                                        [Values(0u)] uint rd,
+                                        [Values(1u)] uint rn,
+                                        [Values(5u)] uint rm,
+                                        [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                        [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                        [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table2,
+                                        [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table3,
+                                        [ValueSource("_GenIdxsForTbl4_")] ulong indexes,
+                                        [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v0 = MakeVectorE0E1(z, z);
+            Vector128<float> v1 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v2 = MakeVectorE0E1(table1, table1);
+            Vector128<float> v3 = MakeVectorE0E1(table2, table2);
+            Vector128<float> v4 = MakeVectorE0E1(table3, table3);
+            Vector128<float> v5 = MakeVectorE0E1(indexes, q == 1u ? indexes : 0ul);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2, v3: v3, v4: v4, v5: v5);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("TBL <Vd>.<Ta>, { <Vn>.16B, <Vn+1>.16B, <Vn+2>.16B, <Vn+3>.16B }, <Vm>.<Ta>")]
+        public void Mod_FourRegTbl_V_8B_16B([ValueSource("_FourRegTbl_V_8B_16B_")] uint opcodes,
+                                            [Values(30u, 3u)] uint rd,
+                                            [Values(31u)]     uint rn,
+                                            [Values(3u, 30u)] uint rm,
+                                            [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table0,
+                                            [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table1,
+                                            [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table2,
+                                            [ValueSource("_8B_")] [Random(RndCntTbls)] ulong table3,
+                                            [ValueSource("_GenIdxsForTbl4_")] ulong indexes,
+                                            [Values(0b0u, 0b1u)] uint q) // <8B, 16B>
+        {
+            opcodes |= ((rm & 31) << 16) | ((rn & 31) << 5) | ((rd & 31) << 0);
+            opcodes |= ((q & 1) << 30);
+
+            ulong z = TestContext.CurrentContext.Random.NextULong();
+            Vector128<float> v30 = MakeVectorE0E1(z, z);
+            Vector128<float> v31 = MakeVectorE0E1(table0, table0);
+            Vector128<float> v0  = MakeVectorE0E1(table1, table1);
+            Vector128<float> v1  = MakeVectorE0E1(table2, table2);
+            Vector128<float> v2  = MakeVectorE0E1(table3, table3);
+            Vector128<float> v3  = MakeVectorE0E1(indexes, indexes);
+
+            SingleOpcode(opcodes, v0: v0, v1: v1, v2: v2, v3: v3, v30: v30, v31: v31);
+
+            CompareAgainstUnicorn();
+        }
+#endif
+    }
+}