ARMeilleure: Respect FZ/RM flags for all floating point operations (#4618)

* ARMeilleure: Respect Fz flag for all floating point operations.

This is a change in strategy for emulating the Fz FPCR flag. Before, it was set before instructions that "needed it" and reset after. However, this missed a few hot instructions like the multiplication instruction, and the entirety of A32.

The new strategy is to set the Fz flag only in the following circumstances:

- Set to match FPCR before translated functions/loop are executed.
- Reset when calling SoftFloat methods, set when returning.
- Reset when exiting execution.

This allows us to remove the code around the existing Fz aware instructions, and get the accuracy benefits on all floating point instructions executed while in translated code.

Single step executions now need to be called with a context wrapper - right now it just contains the Fz flag initialization, and won't actually do anything on ARM.

This fixes a bug in Breath of the Wild where some physics interactions could randomly crash the game due to subnormal values not flushing to zero.

This is draft right now because I need to answer the questions:
- Does dotnet avoid changing the value of Mxcsr?
- Is it a good idea to assume that? Or should the flag set/restore be done on every managed method call, not just softfloat?
- If we assume that, do we want a unit test to verify the behaviour?

I recommend testing a bunch of games, especially games affected when this was originally added, such as #1611.

* Remove unused method

* Use FMA for Fmadd, Fmsub, Fnmadd, Fnmsub, Fmla, Fmls

...when available.

Similar implementation to A32

* Use FMA for Frecps, Frsqrts

* Don't set DAZ.

* Add round mode to ARM FP mode

* Fix mistakes

* Add test for FP state when calling managed methods

* Add explanatory comment to test.

* Cleanup

* Add A64 FPCR flags

* Vrintx_S A32 fast path on A64 backend

* Address feedback 1, re-enable DAZ

* Fix FMA instructions By Elem

* Address feedback

ARMeilleure/CodeGen/Arm64/IntrinsicTable.cs

@@ -226,6 +226,8 @@ namespace ARMeilleure.CodeGen.Arm64
             Add(Intrinsic.Arm64MlsVe,         new IntrinsicInfo(0x2f004000u, IntrinsicType.VectorTernaryRdByElem));
             Add(Intrinsic.Arm64MlsV,          new IntrinsicInfo(0x2e209400u, IntrinsicType.VectorTernaryRd));
             Add(Intrinsic.Arm64MoviV,         new IntrinsicInfo(0x0f000400u, IntrinsicType.VectorMovi));
+            Add(Intrinsic.Arm64MrsFpcr,       new IntrinsicInfo(0xd53b4400u, IntrinsicType.GetRegister));
+            Add(Intrinsic.Arm64MsrFpcr,       new IntrinsicInfo(0xd51b4400u, IntrinsicType.SetRegister));
             Add(Intrinsic.Arm64MrsFpsr,       new IntrinsicInfo(0xd53b4420u, IntrinsicType.GetRegister));
             Add(Intrinsic.Arm64MsrFpsr,       new IntrinsicInfo(0xd51b4420u, IntrinsicType.SetRegister));
             Add(Intrinsic.Arm64MulVe,         new IntrinsicInfo(0x0f008000u, IntrinsicType.VectorBinaryByElem));

ARMeilleure/CodeGen/X86/AssemblerTable.cs

@@ -268,11 +268,13 @@ namespace ARMeilleure.CodeGen.X86
             Add(X86Instruction.Vblendvps,     new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f3a4a, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vcvtph2ps,     new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f3813, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vcvtps2ph,     new InstructionInfo(0x000f3a1d, BadOp,      BadOp,      BadOp,      BadOp,      InstructionFlags.Vex | InstructionFlags.Prefix66));
+            Add(X86Instruction.Vfmadd231pd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b8, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmadd231ps,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b8, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfmadd231sd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b9, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmadd231ss,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b9, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfmsub231sd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bb, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmsub231ss,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bb, InstructionFlags.Vex | InstructionFlags.Prefix66));
+            Add(X86Instruction.Vfnmadd231pd,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bc, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfnmadd231ps,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bc, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfnmadd231sd,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bd, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfnmadd231ss,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bd, InstructionFlags.Vex | InstructionFlags.Prefix66));

ARMeilleure/CodeGen/X86/CodeGenerator.cs

@@ -249,10 +249,9 @@ namespace ARMeilleure.CodeGen.X86
                     case IntrinsicType.Mxcsr:
                     {
                         Operand offset = operation.GetSource(0);
-                        Operand bits   = operation.GetSource(1);
 
-                        Debug.Assert(offset.Kind == OperandKind.Constant && bits.Kind == OperandKind.Constant);
-                        Debug.Assert(offset.Type == OperandType.I32 && bits.Type == OperandType.I32);
+                        Debug.Assert(offset.Kind == OperandKind.Constant);
+                        Debug.Assert(offset.Type == OperandType.I32);
 
                         int offs = offset.AsInt32() + context.CallArgsRegionSize;
 
@@ -261,21 +260,23 @@ namespace ARMeilleure.CodeGen.X86
 
                         Debug.Assert(HardwareCapabilities.SupportsSse || HardwareCapabilities.SupportsVexEncoding);
 
-                        context.Assembler.Stmxcsr(memOp);
-
-                        if (operation.Intrinsic == Intrinsic.X86Mxcsrmb)
+                        if (operation.Intrinsic == Intrinsic.X86Ldmxcsr)
                         {
-                            context.Assembler.Or(memOp, bits, OperandType.I32);
+                            Operand bits = operation.GetSource(1);
+                            Debug.Assert(bits.Type == OperandType.I32);
+
+                            context.Assembler.Mov(memOp, bits, OperandType.I32);
+                            context.Assembler.Ldmxcsr(memOp);
                         }
-                        else /* if (intrinOp.Intrinsic == Intrinsic.X86Mxcsrub) */
+                        else if (operation.Intrinsic == Intrinsic.X86Stmxcsr)
                         {
-                            Operand notBits = Const(~bits.AsInt32());
+                            Operand dest = operation.Destination;
+                            Debug.Assert(dest.Type == OperandType.I32);
 
-                            context.Assembler.And(memOp, notBits, OperandType.I32);
+                            context.Assembler.Stmxcsr(memOp);
+                            context.Assembler.Mov(dest, memOp, OperandType.I32);
                         }
 
-                        context.Assembler.Ldmxcsr(memOp);
-
                         break;
                     }
 

ARMeilleure/CodeGen/X86/IntrinsicTable.cs

@@ -60,6 +60,7 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Haddpd,        new IntrinsicInfo(X86Instruction.Haddpd,        IntrinsicType.Binary));
             Add(Intrinsic.X86Haddps,        new IntrinsicInfo(X86Instruction.Haddps,        IntrinsicType.Binary));
             Add(Intrinsic.X86Insertps,      new IntrinsicInfo(X86Instruction.Insertps,      IntrinsicType.TernaryImm));
+            Add(Intrinsic.X86Ldmxcsr,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr));
             Add(Intrinsic.X86Maxpd,         new IntrinsicInfo(X86Instruction.Maxpd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Maxps,         new IntrinsicInfo(X86Instruction.Maxps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Maxsd,         new IntrinsicInfo(X86Instruction.Maxsd,         IntrinsicType.Binary));
@@ -75,8 +76,6 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Mulps,         new IntrinsicInfo(X86Instruction.Mulps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Mulsd,         new IntrinsicInfo(X86Instruction.Mulsd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Mulss,         new IntrinsicInfo(X86Instruction.Mulss,         IntrinsicType.Binary));
-            Add(Intrinsic.X86Mxcsrmb,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr)); // Mask bits.
-            Add(Intrinsic.X86Mxcsrub,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr)); // Unmask bits.
             Add(Intrinsic.X86Paddb,         new IntrinsicInfo(X86Instruction.Paddb,         IntrinsicType.Binary));
             Add(Intrinsic.X86Paddd,         new IntrinsicInfo(X86Instruction.Paddd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Paddq,         new IntrinsicInfo(X86Instruction.Paddq,         IntrinsicType.Binary));
@@ -160,6 +159,7 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Sqrtps,        new IntrinsicInfo(X86Instruction.Sqrtps,        IntrinsicType.Unary));
             Add(Intrinsic.X86Sqrtsd,        new IntrinsicInfo(X86Instruction.Sqrtsd,        IntrinsicType.Unary));
             Add(Intrinsic.X86Sqrtss,        new IntrinsicInfo(X86Instruction.Sqrtss,        IntrinsicType.Unary));
+            Add(Intrinsic.X86Stmxcsr,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr));
             Add(Intrinsic.X86Subpd,         new IntrinsicInfo(X86Instruction.Subpd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Subps,         new IntrinsicInfo(X86Instruction.Subps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Subsd,         new IntrinsicInfo(X86Instruction.Subsd,         IntrinsicType.Binary));
@@ -170,11 +170,13 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Unpcklps,      new IntrinsicInfo(X86Instruction.Unpcklps,      IntrinsicType.Binary));
             Add(Intrinsic.X86Vcvtph2ps,     new IntrinsicInfo(X86Instruction.Vcvtph2ps,     IntrinsicType.Unary));
             Add(Intrinsic.X86Vcvtps2ph,     new IntrinsicInfo(X86Instruction.Vcvtps2ph,     IntrinsicType.BinaryImm));
+            Add(Intrinsic.X86Vfmadd231pd,   new IntrinsicInfo(X86Instruction.Vfmadd231pd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231ps,   new IntrinsicInfo(X86Instruction.Vfmadd231ps,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231sd,   new IntrinsicInfo(X86Instruction.Vfmadd231sd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231ss,   new IntrinsicInfo(X86Instruction.Vfmadd231ss,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmsub231sd,   new IntrinsicInfo(X86Instruction.Vfmsub231sd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmsub231ss,   new IntrinsicInfo(X86Instruction.Vfmsub231ss,   IntrinsicType.Fma));
+            Add(Intrinsic.X86Vfnmadd231pd,  new IntrinsicInfo(X86Instruction.Vfnmadd231pd,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231ps,  new IntrinsicInfo(X86Instruction.Vfnmadd231ps,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231sd,  new IntrinsicInfo(X86Instruction.Vfnmadd231sd,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231ss,  new IntrinsicInfo(X86Instruction.Vfnmadd231ss,  IntrinsicType.Fma));

ARMeilleure/CodeGen/X86/Mxcsr.cs

@@ -0,0 +1,15 @@
+using System;
+
+namespace ARMeilleure.CodeGen.X86
+{
+    [Flags]
+    enum Mxcsr
+    {
+        Ftz = 1 << 15, // Flush To Zero.
+        Rhi = 1 << 14, // Round Mode high bit.
+        Rlo = 1 << 13, // Round Mode low bit.
+        Um = 1 << 11,  // Underflow Mask.
+        Dm = 1 << 8,   // Denormal Mask.
+        Daz = 1 << 6   // Denormals Are Zero.
+    }
+}

ARMeilleure/CodeGen/X86/PreAllocator.cs

@@ -120,12 +120,18 @@ namespace ARMeilleure.CodeGen.X86
                             break;
 
                         case Instruction.Extended:
-                            if (node.Intrinsic == Intrinsic.X86Mxcsrmb || node.Intrinsic == Intrinsic.X86Mxcsrub)
+                            if (node.Intrinsic == Intrinsic.X86Ldmxcsr)
                             {
                                 int stackOffset = stackAlloc.Allocate(OperandType.I32);
 
                                 node.SetSources(new Operand[] { Const(stackOffset), node.GetSource(0) });
                             }
+                            else if (node.Intrinsic == Intrinsic.X86Stmxcsr)
+                            {
+                                int stackOffset = stackAlloc.Allocate(OperandType.I32);
+
+                                node.SetSources(new Operand[] { Const(stackOffset) });
+                            }
                             break;
                     }
                 }

ARMeilleure/CodeGen/X86/X86Instruction.cs

@@ -208,11 +208,13 @@ namespace ARMeilleure.CodeGen.X86
         Vblendvps,
         Vcvtph2ps,
         Vcvtps2ph,
+        Vfmadd231pd,
         Vfmadd231ps,
         Vfmadd231sd,
         Vfmadd231ss,
         Vfmsub231sd,
         Vfmsub231ss,
+        Vfnmadd231pd,
         Vfnmadd231ps,
         Vfnmadd231sd,
         Vfnmadd231ss,

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -615,14 +615,11 @@ namespace ARMeilleure.Instructions
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                     {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                        {
-                            IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
+                        IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
 
-                            Intrinsic addInst = (op.Size & 1) == 0 ? Intrinsic.X86Addps : Intrinsic.X86Addpd;
+                        Intrinsic addInst = (op.Size & 1) == 0 ? Intrinsic.X86Addps : Intrinsic.X86Addpd;
 
-                            return context.AddIntrinsic(addInst, op1, op2);
-                        }, scalar: false, op1, op2);
+                        return context.AddIntrinsic(addInst, op1, op2);
                     }, scalar: false, op1, op2);
                 });
             }
@@ -696,17 +693,33 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Addss, a, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ss, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addss, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Addsd, a, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231sd, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addsd, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -730,10 +743,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                    }, scalar: true, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                 }, scalar: true);
             }
             else
@@ -755,10 +765,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                    }, scalar: false, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                 }, scalar: false);
             }
             else
@@ -886,10 +893,7 @@ namespace ARMeilleure.Instructions
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                     {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                        {
-                            return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                        }, scalar: false, op1, op2);
+                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                     }, scalar: false, op1, op2);
                 });
             }
@@ -914,10 +918,7 @@ namespace ARMeilleure.Instructions
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                     {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                        {
-                            return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                        }, scalar: false, op1, op2);
+                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                     }, scalar: false, op1, op2);
                 });
             }
@@ -940,10 +941,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                    }, scalar: true, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                 }, scalar: true);
             }
             else
@@ -965,10 +963,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                    }, scalar: false, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                 }, scalar: false);
             }
             else
@@ -1096,10 +1091,7 @@ namespace ARMeilleure.Instructions
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                     {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                        {
-                            return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                        }, scalar: false, op1, op2);
+                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                     }, scalar: false, op1, op2);
                 });
             }
@@ -1124,10 +1116,7 @@ namespace ARMeilleure.Instructions
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                     {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                        {
-                            return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                        }, scalar: false, op1, op2);
+                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                     }, scalar: false, op1, op2);
                 });
             }
@@ -1146,6 +1135,37 @@ namespace ARMeilleure.Instructions
             {
                 InstEmitSimdHelperArm64.EmitScalarTernaryOpFRdByElem(context, Intrinsic.Arm64FmlaSe);
             }
+            else if (Optimizations.UseFma)
+            {
+                OpCodeSimdRegElemF op = (OpCodeSimdRegElemF)context.CurrOp;
+
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+                Operand m = GetVec(op.Rm);
+
+                int sizeF = op.Size & 1;
+
+                if (sizeF == 0)
+                {
+                    int shuffleMask = op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ss, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper96(res));
+                }
+                else /* if (sizeF == 1) */
+                {
+                    int shuffleMask = op.Index | op.Index << 1;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfmadd231sd, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper64(res));
+                }
+            }
             else
             {
                 EmitScalarTernaryOpByElemF(context, (op1, op2, op3) =>
@@ -1171,11 +1191,19 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
-
-                    res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ps, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1186,9 +1214,15 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (sizeF == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
-
-                    res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231pd, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1224,8 +1258,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
 
-                    res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ps, d, n, res);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1240,8 +1281,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
 
-                    res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231pd, d, n, res);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1261,6 +1309,37 @@ namespace ARMeilleure.Instructions
             {
                 InstEmitSimdHelperArm64.EmitScalarTernaryOpFRdByElem(context, Intrinsic.Arm64FmlsSe);
             }
+            else if (Optimizations.UseFma)
+            {
+                OpCodeSimdRegElemF op = (OpCodeSimdRegElemF)context.CurrOp;
+
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+                Operand m = GetVec(op.Rm);
+
+                int sizeF = op.Size & 1;
+
+                if (sizeF == 0)
+                {
+                    int shuffleMask = op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper96(res));
+                }
+                else /* if (sizeF == 1) */
+                {
+                    int shuffleMask = op.Index | op.Index << 1;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper64(res));
+                }
+            }
             else
             {
                 EmitScalarTernaryOpByElemF(context, (op1, op2, op3) =>
@@ -1286,11 +1365,19 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
-
-                    res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1301,9 +1388,15 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (sizeF == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
-
-                    res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1339,8 +1432,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
 
-                    res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, d, n, res);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1355,8 +1455,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
 
-                    res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, d, n, res);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1385,17 +1492,33 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Subss, a, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subss, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Subsd, a, res);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subsd, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1669,25 +1792,39 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand mask = X86GetScalar(context, -0f);
-
-                    Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmsub231ss, a, n, m);
+                    }
+                    else
+                    {
+                        Operand mask = X86GetScalar(context, -0f);
+                        Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Subss, aNeg, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subss, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand mask = X86GetScalar(context, -0d);
-
-                    Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmsub231sd, a, n, m);
+                    }
+                    else
+                    {
+                        Operand mask = X86GetScalar(context, -0d);
+                        Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Subsd, aNeg, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subsd, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1716,25 +1853,39 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand mask = X86GetScalar(context, -0f);
-
-                    Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmsub231ss, a, n, m);
+                    }
+                    else
+                    {
+                        Operand mask = X86GetScalar(context, -0f);
+                        Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Addss, aNeg, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addss, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand mask = X86GetScalar(context, -0d);
-
-                    Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmsub231sd, a, n, m);
+                    }
+                    else
+                    {
+                        Operand mask = X86GetScalar(context, -0d);
+                        Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
-                            res = context.AddIntrinsic(Intrinsic.X86Addsd, aNeg, res);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Addsd, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1830,13 +1981,22 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand mask = X86GetScalar(context, 2f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subss, mask, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subss, mask, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper96(res));
@@ -1845,9 +2005,16 @@ namespace ARMeilleure.Instructions
                 {
                     Operand mask = X86GetScalar(context, 2d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subsd, mask, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subsd, mask, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper64(res));
@@ -1877,14 +2044,23 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand mask = X86GetAllElements(context, 2f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
-                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subps, mask, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subps, mask, res);
+                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1897,10 +2073,17 @@ namespace ARMeilleure.Instructions
                 {
                     Operand mask = X86GetAllElements(context, 2d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
-                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subpd, mask, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subpd, mask, res);
+                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
 
                     context.Copy(GetVec(op.Rd), res);
                 }
@@ -2113,20 +2296,32 @@ namespace ARMeilleure.Instructions
 
         public static void Frintx_S(ArmEmitterContext context)
         {
-            // TODO Arm64: Fast path. Should we set host FPCR?
-            EmitScalarUnaryOpF(context, (op1) =>
+            if (Optimizations.UseAdvSimd)
+            {
+                InstEmitSimdHelperArm64.EmitScalarUnaryOpF(context, Intrinsic.Arm64FrintxS);
+            }
+            else
             {
-                return EmitRoundByRMode(context, op1);
-            });
+                EmitScalarUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundByRMode(context, op1);
+                });
+            }
         }
 
         public static void Frintx_V(ArmEmitterContext context)
         {
-            // TODO Arm64: Fast path. Should we set host FPCR?
-            EmitVectorUnaryOpF(context, (op1) =>
+            if (Optimizations.UseAdvSimd)
+            {
+                InstEmitSimdHelperArm64.EmitVectorUnaryOpF(context, Intrinsic.Arm64FrintxV);
+            }
+            else
             {
-                return EmitRoundByRMode(context, op1);
-            });
+                EmitVectorUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundByRMode(context, op1);
+                });
+            }
         }
 
         public static void Frintz_S(ArmEmitterContext context)
@@ -2237,16 +2432,25 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand maskHalf    = X86GetScalar(context, 0.5f);
                     Operand maskThree   = X86GetScalar(context, 3f);
                     Operand maskOneHalf = X86GetScalar(context, 1.5f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subss, maskThree, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subss, maskThree, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Mulss, maskHalf,  res);
+                    res = context.AddIntrinsic(Intrinsic.X86Mulss, maskHalf, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper96(res));
@@ -2257,10 +2461,17 @@ namespace ARMeilleure.Instructions
                     Operand maskThree   = X86GetScalar(context, 3d);
                     Operand maskOneHalf = X86GetScalar(context, 1.5d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subsd, maskThree, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subsd, maskThree, res);
-                    res = context.AddIntrinsic(Intrinsic.X86Mulsd, maskHalf,  res);
+                    res = context.AddIntrinsic(Intrinsic.X86Mulsd, maskHalf, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper64(res));
@@ -2290,15 +2501,24 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand maskHalf    = X86GetAllElements(context, 0.5f);
                     Operand maskThree   = X86GetAllElements(context, 3f);
                     Operand maskOneHalf = X86GetAllElements(context, 1.5f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subps, maskThree, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subps, maskThree, res);
                     res = context.AddIntrinsic(Intrinsic.X86Mulps, maskHalf,  res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: false, sizeF);
 
@@ -2315,9 +2535,16 @@ namespace ARMeilleure.Instructions
                     Operand maskThree   = X86GetAllElements(context, 3d);
                     Operand maskOneHalf = X86GetAllElements(context, 1.5d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Subpd, maskThree, res);
+                    }
 
-                    res = context.AddIntrinsic(Intrinsic.X86Subpd, maskThree, res);
                     res = context.AddIntrinsic(Intrinsic.X86Mulpd, maskHalf,  res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: false, sizeF);
 
@@ -4728,53 +4955,6 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        public static Operand EmitSseOrAvxHandleFzModeOpF(
-            ArmEmitterContext context,
-            Func2I emit,
-            bool scalar,
-            Operand n = default,
-            Operand m = default)
-        {
-            Operand nCopy = n == default ? context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rn)) : n;
-            Operand mCopy = m == default ? context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rm)) : m;
-
-            EmitSseOrAvxEnterFtzAndDazModesOpF(context, out Operand isTrue);
-
-            Operand res = emit(nCopy, mCopy);
-
-            EmitSseOrAvxExitFtzAndDazModesOpF(context, isTrue);
-
-            if (n != default || m != default)
-            {
-                return res;
-            }
-
-            int sizeF = ((IOpCodeSimd)context.CurrOp).Size & 1;
-
-            if (sizeF == 0)
-            {
-                if (scalar)
-                {
-                    res = context.VectorZeroUpper96(res);
-                }
-                else if (((OpCodeSimdReg)context.CurrOp).RegisterSize == RegisterSize.Simd64)
-                {
-                    res = context.VectorZeroUpper64(res);
-                }
-            }
-            else /* if (sizeF == 1) */
-            {
-                if (scalar)
-                {
-                    res = context.VectorZeroUpper64(res);
-                }
-            }
-
-            context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rd), res);
-
-            return default;
-        }
-
         private static Operand EmitSse2VectorMaxMinOpF(ArmEmitterContext context, Operand n, Operand m, bool isMax)
         {
             IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
@@ -4834,10 +5014,7 @@ namespace ARMeilleure.Instructions
 
                 Operand res = EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
-                    }, scalar: scalar, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
                 }, scalar: scalar, nCopy, mCopy);
 
                 if (n != default || m != default)
@@ -4872,10 +5049,7 @@ namespace ARMeilleure.Instructions
 
                 Operand res = EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
                 {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
-                    {
-                        return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
-                    }, scalar: scalar, op1, op2);
+                    return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
                 }, scalar: scalar, nCopy, mCopy);
 
                 if (n != default || m != default)

ARMeilleure/Instructions/InstEmitSimdCvt32.cs

@@ -356,9 +356,11 @@ namespace ARMeilleure.Instructions
                         ? typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert))
                         : typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert));
 
+                    context.ExitArmFpMode();
                     context.StoreToContext();
                     Operand res = context.Call(method, src);
                     context.LoadFromContext();
+                    context.EnterArmFpMode();
 
                     InsertScalar16(context, op.Vd, op.T, res);
                 }
@@ -372,9 +374,11 @@ namespace ARMeilleure.Instructions
                         ? typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert))
                         : typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert));
 
+                    context.ExitArmFpMode();
                     context.StoreToContext();
                     Operand res = context.Call(method, src);
                     context.LoadFromContext();
+                    context.EnterArmFpMode();
 
                     InsertScalar(context, op.Vd, res);
                 }
@@ -542,10 +546,17 @@ namespace ARMeilleure.Instructions
         // VRINTX (floating-point).
         public static void Vrintx_S(ArmEmitterContext context)
         {
-            EmitScalarUnaryOpF32(context, (op1) =>
+            if (Optimizations.UseAdvSimd)
             {
-                return EmitRoundByRMode(context, op1);
-            });
+                InstEmitSimdHelper32Arm64.EmitScalarUnaryOpF32(context, Intrinsic.Arm64FrintxS);
+            }
+            else
+            {
+                EmitScalarUnaryOpF32(context, (op1) =>
+                {
+                    return EmitRoundByRMode(context, op1);
+                });
+            }
         }
 
         private static Operand EmitFPConvert(ArmEmitterContext context, Operand value, OperandType type, bool signed)

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -1,3 +1,4 @@
+using ARMeilleure.CodeGen.X86;
 using ARMeilleure.Decoders;
 using ARMeilleure.IntermediateRepresentation;
 using ARMeilleure.State;
@@ -158,6 +159,75 @@ namespace ARMeilleure.Instructions
         };
 #endregion
 
+        public static void EnterArmFpMode(EmitterContext context, Func<FPState, Operand> getFpFlag)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                Operand fzTrue = getFpFlag(FPState.FzFlag);
+                Operand r0True = getFpFlag(FPState.RMode0Flag);
+                Operand r1True = getFpFlag(FPState.RMode1Flag);
+
+                mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Rhi | Mxcsr.Rlo)));
+
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(fzTrue, Const((int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Um | Mxcsr.Dm)), Const(0)));
+
+                // X86 round modes in order: nearest, negative, positive, zero
+                // ARM round modes in order: nearest, positive, negative, zero
+                // Read the bits backwards to correct this.
+
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(r0True, Const((int)Mxcsr.Rhi), Const(0)));
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(r1True, Const((int)Mxcsr.Rlo), Const(0)));
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                Operand fzTrue = getFpFlag(FPState.FzFlag);
+                Operand r0True = getFpFlag(FPState.RMode0Flag);
+                Operand r1True = getFpFlag(FPState.RMode1Flag);
+
+                fpcr = context.BitwiseAnd(fpcr, Const(~(int)(FPCR.Fz | FPCR.RMode0 | FPCR.RMode1)));
+
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(fzTrue, Const((int)FPCR.Fz), Const(0)));
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(r0True, Const((int)FPCR.RMode0), Const(0)));
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(r1True, Const((int)FPCR.RMode1), Const(0)));
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                // TODO: Restore FPSR
+            }
+        }
+
+        public static void ExitArmFpMode(EmitterContext context, Action<FPState, Operand> setFpFlag)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                // Unset round mode (to nearest) and ftz.
+                mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Rhi | Mxcsr.Rlo)));
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+
+                // Status flags would be stored here if they were used.
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                // Unset round mode (to nearest) and fz.
+                fpcr = context.BitwiseAnd(fpcr, Const(~(int)(FPCR.Fz | FPCR.RMode0 | FPCR.RMode1)));
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                // TODO: Store FPSR
+            }
+        }
+
         public static int GetImmShl(OpCodeSimdShImm op)
         {
             return op.Imm - (8 << op.Size);
@@ -465,9 +535,11 @@ namespace ARMeilleure.Instructions
                 ? typeof(SoftFloat32).GetMethod(name)
                 : typeof(SoftFloat64).GetMethod(name);
 
+            context.ExitArmFpMode();
             context.StoreToContext();
             Operand res = context.Call(info, callArgs);
             context.LoadFromContext();
+            context.EnterArmFpMode();
 
             return res;
         }
@@ -1358,39 +1430,6 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        [Flags]
-        public enum Mxcsr
-        {
-            Ftz = 1 << 15, // Flush To Zero.
-            Um  = 1 << 11, // Underflow Mask.
-            Dm  = 1 << 8,  // Denormal Mask.
-            Daz = 1 << 6   // Denormals Are Zero.
-        }
-
-        public static void EmitSseOrAvxEnterFtzAndDazModesOpF(ArmEmitterContext context, out Operand isTrue)
-        {
-            isTrue = GetFpFlag(FPState.FzFlag);
-
-            Operand lblTrue = Label();
-            context.BranchIfFalse(lblTrue, isTrue);
-
-            context.AddIntrinsicNoRet(Intrinsic.X86Mxcsrmb, Const((int)(Mxcsr.Ftz | Mxcsr.Um | Mxcsr.Dm | Mxcsr.Daz)));
-
-            context.MarkLabel(lblTrue);
-        }
-
-        public static void EmitSseOrAvxExitFtzAndDazModesOpF(ArmEmitterContext context, Operand isTrue = default)
-        {
-            isTrue = isTrue == default ? GetFpFlag(FPState.FzFlag) : isTrue;
-
-            Operand lblTrue = Label();
-            context.BranchIfFalse(lblTrue, isTrue);
-
-            context.AddIntrinsicNoRet(Intrinsic.X86Mxcsrub, Const((int)(Mxcsr.Ftz | Mxcsr.Daz)));
-
-            context.MarkLabel(lblTrue);
-        }
-
         public enum CmpCondition
         {
             // Legacy Sse.

ARMeilleure/Instructions/InstEmitSimdHelper32.cs

@@ -1197,9 +1197,11 @@ namespace ARMeilleure.Instructions
             Array.Resize(ref callArgs, callArgs.Length + 1);
             callArgs[callArgs.Length - 1] = Const(1);
 
+            context.ExitArmFpMode();
             context.StoreToContext();
             Operand res = context.Call(info, callArgs);
             context.LoadFromContext();
+            context.EnterArmFpMode();
 
             return res;
         }

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -192,6 +192,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpcr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
 
         private static void EmitSetFpsr(ArmEmitterContext context)
@@ -210,6 +212,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpsr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
     }
 }

ARMeilleure/Instructions/InstEmitSystem32.cs

@@ -321,6 +321,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpscr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
     }
 }

ARMeilleure/IntermediateRepresentation/Intrinsic.cs

@@ -53,6 +53,7 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Haddpd,
         X86Haddps,
         X86Insertps,
+        X86Ldmxcsr,
         X86Maxpd,
         X86Maxps,
         X86Maxsd,
@@ -68,8 +69,6 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Mulps,
         X86Mulsd,
         X86Mulss,
-        X86Mxcsrmb,
-        X86Mxcsrub,
         X86Paddb,
         X86Paddd,
         X86Paddq,
@@ -153,6 +152,7 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Sqrtps,
         X86Sqrtsd,
         X86Sqrtss,
+        X86Stmxcsr,
         X86Subpd,
         X86Subps,
         X86Subsd,
@@ -163,11 +163,13 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Unpcklps,
         X86Vcvtph2ps,
         X86Vcvtps2ph,
+        X86Vfmadd231pd,
         X86Vfmadd231ps,
         X86Vfmadd231sd,
         X86Vfmadd231ss,
         X86Vfmsub231sd,
         X86Vfmsub231ss,
+        X86Vfnmadd231pd,
         X86Vfnmadd231ps,
         X86Vfnmadd231sd,
         X86Vfnmadd231ss,
@@ -394,6 +396,8 @@ namespace ARMeilleure.IntermediateRepresentation
         Arm64MlsVe,
         Arm64MlsV,
         Arm64MoviV,
+        Arm64MrsFpcr,
+        Arm64MsrFpcr,
         Arm64MrsFpsr,
         Arm64MsrFpsr,
         Arm64MulVe,

ARMeilleure/Translation/ArmEmitterContext.cs

@@ -188,6 +188,21 @@ namespace ARMeilleure.Translation
             }
         }
 
+        public void EnterArmFpMode()
+        {
+            InstEmitSimdHelper.EnterArmFpMode(this, InstEmitHelper.GetFpFlag);
+        }
+
+        public void UpdateArmFpMode()
+        {
+            EnterArmFpMode();
+        }
+
+        public void ExitArmFpMode()
+        {
+            InstEmitSimdHelper.ExitArmFpMode(this, (flag, value) => InstEmitHelper.SetFpFlag(this, flag, value));
+        }
+
         public Operand TryGetComparisonResult(Condition condition)
         {
             if (_optOpLastCompare == null || _optOpLastCompare != _optOpLastFlagSet)

ARMeilleure/Translation/DispatcherFunction.cs

@@ -3,4 +3,5 @@
 namespace ARMeilleure.Translation
 {
     delegate void DispatcherFunction(IntPtr nativeContext, ulong startAddress);
+    delegate ulong WrapperFunction(IntPtr nativeContext, ulong startAddress);
 }

ARMeilleure/Translation/PTC/Ptc.cs

@@ -30,7 +30,7 @@ namespace ARMeilleure.Translation.PTC
         private const string OuterHeaderMagicString = "PTCohd\0\0";
         private const string InnerHeaderMagicString = "PTCihd\0\0";
 
-        private const uint InternalVersion = 4485; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 4626; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";

ARMeilleure/Translation/TranslatedFunction.cs

@@ -25,5 +25,10 @@ namespace ARMeilleure.Translation
         {
             return _func(context.NativeContextPtr);
         }
+
+        public ulong Execute(WrapperFunction dispatcher, State.ExecutionContext context)
+        {
+            return dispatcher(context.NativeContextPtr, (ulong)FuncPointer);
+        }
     }
 }

ARMeilleure/Translation/Translator.cs

@@ -183,7 +183,7 @@ namespace ARMeilleure.Translation
 
             Statistics.StartTimer();
 
-            ulong nextAddr = func.Execute(context);
+            ulong nextAddr = func.Execute(Stubs.ContextWrapper, context);
 
             Statistics.StopTimer(address);
 
@@ -194,7 +194,7 @@ namespace ARMeilleure.Translation
         {
             TranslatedFunction func = Translate(address, context.ExecutionMode, highCq: false, singleStep: true);
 
-            address = func.Execute(context);
+            address = func.Execute(Stubs.ContextWrapper, context);
 
             EnqueueForDeletion(address, func);
 

ARMeilleure/Translation/TranslatorStubs.cs

@@ -21,6 +21,7 @@ namespace ARMeilleure.Translation
         private readonly Translator _translator;
         private readonly Lazy<IntPtr> _dispatchStub;
         private readonly Lazy<DispatcherFunction> _dispatchLoop;
+        private readonly Lazy<WrapperFunction> _contextWrapper;
 
         /// <summary>
         /// Gets the dispatch stub.
@@ -64,6 +65,20 @@ namespace ARMeilleure.Translation
             }
         }
 
+        /// <summary>
+        /// Gets the context wrapper function.
+        /// </summary>
+        /// <exception cref="ObjectDisposedException"><see cref="TranslatorStubs"/> instance was disposed</exception>
+        public WrapperFunction ContextWrapper
+        {
+            get
+            {
+                ObjectDisposedException.ThrowIf(_disposed, this);
+
+                return _contextWrapper.Value;
+            }
+        }
+
         /// <summary>
         /// Initializes a new instance of the <see cref="TranslatorStubs"/> class with the specified
         /// <see cref="Translator"/> instance.
@@ -77,6 +92,7 @@ namespace ARMeilleure.Translation
             _translator = translator;
             _dispatchStub = new(GenerateDispatchStub, isThreadSafe: true);
             _dispatchLoop = new(GenerateDispatchLoop, isThreadSafe: true);
+            _contextWrapper = new(GenerateContextWrapper, isThreadSafe: true);
         }
 
         /// <summary>
@@ -202,6 +218,32 @@ namespace ARMeilleure.Translation
             return Marshal.GetFunctionPointerForDelegate(func);
         }
 
+        /// <summary>
+        /// Emits code that syncs FP state before executing guest code, or returns it to normal.
+        /// </summary>
+        /// <param name="context">Emitter context for the method</param>
+        /// <param name="nativeContext">Pointer to the native context</param>
+        /// <param name="enter">True if entering guest code, false otherwise</param>
+        private void EmitSyncFpContext(EmitterContext context, Operand nativeContext, bool enter)
+        {
+            if (enter)
+            {
+                InstEmitSimdHelper.EnterArmFpMode(context, (flag) =>
+                {
+                    Operand flagAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRegisterOffset(new Register((int)flag, RegisterType.FpFlag))));
+                    return context.Load(OperandType.I32, flagAddress);
+                });
+            }
+            else
+            {
+                InstEmitSimdHelper.ExitArmFpMode(context, (flag, value) =>
+                {
+                    Operand flagAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRegisterOffset(new Register((int)flag, RegisterType.FpFlag))));
+                    context.Store(flagAddress, value);
+                });
+            }
+        }
+
         /// <summary>
         /// Generates a <see cref="DispatchLoop"/> function.
         /// </summary>
@@ -221,6 +263,8 @@ namespace ARMeilleure.Translation
             Operand runningAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRunningOffset()));
             Operand dispatchAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetDispatchAddressOffset()));
 
+            EmitSyncFpContext(context, nativeContext, true);
+
             context.MarkLabel(beginLbl);
             context.Store(dispatchAddress, guestAddress);
             context.Copy(guestAddress, context.Call(Const((ulong)DispatchStub), OperandType.I64, nativeContext));
@@ -229,6 +273,9 @@ namespace ARMeilleure.Translation
             context.Branch(beginLbl);
 
             context.MarkLabel(endLbl);
+
+            EmitSyncFpContext(context, nativeContext, false);
+
             context.Return();
 
             var cfg = context.GetControlFlowGraph();
@@ -237,5 +284,29 @@ namespace ARMeilleure.Translation
 
             return Compiler.Compile(cfg, argTypes, retType, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<DispatcherFunction>();
         }
+
+        /// <summary>
+        /// Generates a <see cref="ContextWrapper"/> function.
+        /// </summary>
+        /// <returns><see cref="ContextWrapper"/> function</returns>
+        private WrapperFunction GenerateContextWrapper()
+        {
+            var context = new EmitterContext();
+
+            Operand nativeContext = context.LoadArgument(OperandType.I64, 0);
+            Operand guestMethod = context.LoadArgument(OperandType.I64, 1);
+
+            EmitSyncFpContext(context, nativeContext, true);
+            Operand returnValue = context.Call(guestMethod, OperandType.I64, nativeContext);
+            EmitSyncFpContext(context, nativeContext, false);
+
+            context.Return(returnValue);
+
+            var cfg = context.GetControlFlowGraph();
+            var retType = OperandType.I64;
+            var argTypes = new[] { OperandType.I64, OperandType.I64 };
+
+            return Compiler.Compile(cfg, argTypes, retType, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<WrapperFunction>();
+        }
     }
 }

ARMeilleure/Translation/TranslatorTestMethods.cs

@@ -0,0 +1,148 @@
+using ARMeilleure.CodeGen.X86;
+using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.State;
+using ARMeilleure.Translation;
+using System;
+using System.Runtime.InteropServices;
+using static ARMeilleure.IntermediateRepresentation.Operand.Factory;
+
+namespace ARMeilleure.Translation
+{
+    public static class TranslatorTestMethods
+    {
+        public delegate int FpFlagsPInvokeTest(IntPtr managedMethod);
+
+        private static bool SetPlatformFtz(EmitterContext context, bool ftz)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                if (ftz)
+                {
+                    mxcsr = context.BitwiseOr(mxcsr, Const((int)(Mxcsr.Ftz | Mxcsr.Um | Mxcsr.Dm)));
+                }
+                else
+                {
+                    mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)Mxcsr.Ftz));
+                }
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+
+                return true;
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                if (ftz)
+                {
+                    fpcr = context.BitwiseOr(fpcr, Const((int)FPCR.Fz));
+                }
+                else
+                {
+                    fpcr = context.BitwiseAnd(fpcr, Const(~(int)FPCR.Fz));
+                }
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                return true;
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        private static Operand FpBitsToInt(EmitterContext context, Operand fp)
+        {
+            Operand vec = context.VectorInsert(context.VectorZero(), fp, 0);
+            return context.VectorExtract(OperandType.I32, vec, 0);
+        }
+
+        public static FpFlagsPInvokeTest GenerateFpFlagsPInvokeTest()
+        {
+            EmitterContext context = new EmitterContext();
+
+            Operand methodAddress = context.Copy(context.LoadArgument(OperandType.I64, 0));
+
+            // Verify that default dotnet fp state does not flush to zero.
+            // This is required for SoftFloat to function.
+
+            // Denormal + zero != 0
+
+            Operand denormal = ConstF(BitConverter.Int32BitsToSingle(1)); // 1.40129846432e-45
+            Operand zeroF = ConstF(0f);
+            Operand zero = Const(0);
+
+            Operand result = context.Add(zeroF, denormal);
+
+            // Must not be zero.
+
+            Operand correct1Label = Label();
+
+            context.BranchIfFalse(correct1Label, context.ICompareEqual(FpBitsToInt(context, result), zero));
+
+            context.Return(Const(1));
+
+            context.MarkLabel(correct1Label);
+
+            // Set flush to zero flag. If unsupported by the backend, just return true.
+
+            if (!SetPlatformFtz(context, true))
+            {
+                context.Return(Const(0));
+            }
+
+            // Denormal + zero == 0
+
+            Operand resultFz = context.Add(zeroF, denormal);
+
+            // Must equal zero.
+
+            Operand correct2Label = Label();
+
+            context.BranchIfTrue(correct2Label, context.ICompareEqual(FpBitsToInt(context, resultFz), zero));
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(2));
+
+            context.MarkLabel(correct2Label);
+
+            // Call a managed method. This method should not change Fz state. 
+
+            context.Call(methodAddress, OperandType.None);
+
+            // Denormal + zero == 0
+
+            Operand resultFz2 = context.Add(zeroF, denormal);
+
+            // Must equal zero.
+
+            Operand correct3Label = Label();
+
+            context.BranchIfTrue(correct3Label, context.ICompareEqual(FpBitsToInt(context, resultFz2), zero));
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(3));
+
+            context.MarkLabel(correct3Label);
+
+            // Success.
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(0));
+
+            // Compile and return the function.
+
+            ControlFlowGraph cfg = context.GetControlFlowGraph();
+
+            OperandType[] argTypes = new OperandType[] { OperandType.I64 };
+
+            return Compiler.Compile(cfg, argTypes, OperandType.I32, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<FpFlagsPInvokeTest>();
+        }
+    }
+}

Ryujinx.Tests/Cpu/EnvironmentTests.cs

@@ -0,0 +1,91 @@
+using ARMeilleure.Translation;
+using NUnit.Framework;
+using Ryujinx.Cpu.Jit;
+using Ryujinx.Tests.Memory;
+using System;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Tests.Cpu
+{
+    internal class EnvironmentTests
+    {
+        private static Translator _translator;
+
+        private void EnsureTranslator()
+        {
+            // Create a translator, as one is needed to register the signal handler or emit methods.
+            _translator ??= new Translator(new JitMemoryAllocator(), new MockMemoryManager(), true);
+        }
+
+        [MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.NoOptimization)]
+        private float GetDenormal()
+        {
+            return BitConverter.Int32BitsToSingle(1);
+        }
+
+        [MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.NoOptimization)]
+        private float GetZero()
+        {
+            return BitConverter.Int32BitsToSingle(0);
+        }
+
+        /// <summary>
+        /// This test ensures that managed methods do not reset floating point control flags.
+        /// This is used to avoid changing control flags when running methods that don't require it, such as SVC calls, software memory...
+        /// </summary>
+        [Test]
+        public void FpFlagsPInvoke()
+        {
+            EnsureTranslator();
+
+            // Subnormal results are not flushed to zero by default.
+            // This operation should not be allowed to do constant propagation, hence the methods that explicitly disallow inlining.
+            Assert.AreNotEqual(GetDenormal() + GetZero(), 0f);
+
+            bool methodCalled = false;
+            bool isFz = false;
+
+            var managedMethod = () =>
+            {
+                // Floating point math should not modify fp flags.
+                float test = 2f * 3.5f;
+
+                if (test < 4f)
+                {
+                    throw new System.Exception("Sanity check.");
+                }
+
+                isFz = GetDenormal() + GetZero() == 0f;
+
+                try
+                {
+                    if (test >= 4f)
+                    {
+                        throw new System.Exception("Always throws.");
+                    }
+                }
+                catch
+                {
+                    // Exception handling should not modify fp flags.
+
+                    methodCalled = true;
+                }
+            };
+
+            var method = TranslatorTestMethods.GenerateFpFlagsPInvokeTest();
+
+            // This method sets flush-to-zero and then calls the managed method.
+            // Before and after setting the flags, it ensures subnormal addition works as expected.
+            // It returns a positive result if any tests fail, and 0 on success (or if the platform cannot change FP flags)
+            int result = method(Marshal.GetFunctionPointerForDelegate(managedMethod));
+
+            // Subnormal results are not flushed to zero by default, which we should have returned to exiting the method.
+            Assert.AreNotEqual(GetDenormal() + GetZero(), 0f);
+
+            Assert.True(result == 0);
+            Assert.True(methodCalled);
+            Assert.True(isFz);
+        }
+    }
+}