mirrors/Ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix float constant passed to functions, save and restore non-volatile XMM registers, other fixes

Browse source
This commit is contained in:
gdkchan 2019-06-19 16:38:17 -03:00
parent 87238b509c
commit 0bacc3fd63
7 changed files with 117 additions and 53 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
ARMeilleure/CodeGen/X86/CodeGenContext.cs
View file

@ -21,6 +21,8 @@ namespace ARMeilleure.CodeGen.X86
public int CallArgsRegionSize { get; }
public int VecCalleeSaveSize { get; }
private struct Jump
{
public bool IsConditional { get; }
@ -77,22 +79,27 @@ namespace ARMeilleure.CodeGen.X86
Assembler = new Assembler(stream);
CallArgsRegionSize = GetCallArgsRegionSize(allocResult);
CallArgsRegionSize = GetCallArgsRegionSize(allocResult, out int vecCalleeSaveSize);
VecCalleeSaveSize = vecCalleeSaveSize;
_blockOffsets = new long[blocksCount];
_jumps = new List<Jump>();
}
private int GetCallArgsRegionSize(AllocationResult allocResult)
private int GetCallArgsRegionSize(AllocationResult allocResult, out int vecCalleeSaveSize)
{
//We need to add 8 bytes to the total size, as the call to this
//function already pushed 8 bytes (the return address).
int mask = CallingConvention.GetIntCalleeSavedRegisters() & allocResult.IntUsedRegisters;
int intMask = CallingConvention.GetIntCalleeSavedRegisters() & allocResult.IntUsedRegisters;
int vecMask = CallingConvention.GetVecCalleeSavedRegisters() & allocResult.VecUsedRegisters;
mask |= 1 << (int)X86Register.Rbp;
vecCalleeSaveSize = BitUtils.CountBits(vecMask) * 16;
int calleeSaveRegionSize = CountBits(mask) * 8 + 8;
intMask |= 1 << (int)X86Register.Rbp;
int calleeSaveRegionSize = BitUtils.CountBits(intMask) * 8 + vecCalleeSaveSize + 8;
int argsCount = allocResult.MaxCallArgs;
@ -113,20 +120,6 @@ namespace ARMeilleure.CodeGen.X86
return callArgsAndFrameSize - frameSize;
}
private static int CountBits(int mask)
{
int count = 0;
while (mask != 0)
{
mask &= ~(1 << BitUtils.LowestBitSet(mask));
count++;
}
return count;
}
public void EnterBlock(BasicBlock block)
{
_blockOffsets[block.Index] = _stream.Position;

51
ARMeilleure/CodeGen/X86/CodeGenerator.cs
View file

@ -1793,8 +1793,27 @@ namespace ARMeilleure.CodeGen.X86
mask &= ~(1 << bit);
}
mask = CallingConvention.GetVecCalleeSavedRegisters() & context.AllocResult.VecUsedRegisters;
int offset = 0;
while (mask != 0)
{
int bit = BitUtils.LowestBitSet(mask);
offset -= 16;
X86MemoryOperand memOp = new X86MemoryOperand(OperandType.V128, Register(X86Register.Rsp), null, Scale.x1, offset);
context.Assembler.Movdqu(memOp, Xmm((X86Register)bit));
mask &= ~(1 << bit);
}
int reservedStackSize = context.CallArgsRegionSize + context.AllocResult.SpillRegionSize;
reservedStackSize += context.VecCalleeSaveSize;
if (reservedStackSize != 0)
{
context.Assembler.Sub(Register(X86Register.Rsp), new Operand(reservedStackSize));
@ -1803,17 +1822,36 @@ namespace ARMeilleure.CodeGen.X86
private static void WriteEpilogue(CodeGenContext context)
{
int mask = CallingConvention.GetIntCalleeSavedRegisters() & context.AllocResult.IntUsedRegisters;
mask |= 1 << (int)X86Register.Rbp;
int reservedStackSize = context.CallArgsRegionSize + context.AllocResult.SpillRegionSize;
reservedStackSize += context.VecCalleeSaveSize;
if (reservedStackSize != 0)
{
context.Assembler.Add(Register(X86Register.Rsp), new Operand(reservedStackSize));
}
int mask = CallingConvention.GetVecCalleeSavedRegisters() & context.AllocResult.VecUsedRegisters;
int offset = 0;
while (mask != 0)
{
int bit = BitUtils.LowestBitSet(mask);
offset -= 16;
X86MemoryOperand memOp = new X86MemoryOperand(OperandType.V128, Register(X86Register.Rsp), null, Scale.x1, offset);
context.Assembler.Movdqu(Xmm((X86Register)bit), memOp);
mask &= ~(1 << bit);
}
mask = CallingConvention.GetIntCalleeSavedRegisters() & context.AllocResult.IntUsedRegisters;
mask |= 1 << (int)X86Register.Rbp;
while (mask != 0)
{
int bit = BitUtils.HighestBitSet(mask);
@ -1849,5 +1887,10 @@ namespace ARMeilleure.CodeGen.X86
{
return new Operand((int)register, RegisterType.Integer, type);
}
private static Operand Xmm(X86Register register)
{
return new Operand((int)register, RegisterType.Vector, OperandType.V128);
}
}
}

18
ARMeilleure/CodeGen/X86/PreAllocator.cs
View file

@ -27,7 +27,7 @@ namespace ARMeilleure.CodeGen.X86
Instruction inst = operation.Inst;
AddConstantCopy(node, operation);
HandleConstantCopy(node, operation);
//Comparison instructions uses CMOVcc, which does not zero the
//upper bits of the register (since it's R8), we need to ensure it
@ -55,14 +55,14 @@ namespace ARMeilleure.CodeGen.X86
ReplaceNegateWithXor(node, operation);
}
AddFixedRegisterCopy(node, operation);
HandleFixedRegisterCopy(node, operation);
AddSameDestSrc1Copy(node, operation);
HandleSameDestSrc1Copy(node, operation);
}
}
}
private static void AddConstantCopy(LinkedListNode<Node> node, Operation operation)
private static void HandleConstantCopy(LinkedListNode<Node> node, Operation operation)
{
if (operation.SourcesCount == 0 || IsIntrinsic(operation.Inst))
{
@ -229,7 +229,7 @@ namespace ARMeilleure.CodeGen.X86
Delete(node, operation);
}
private static void AddFixedRegisterCopy(LinkedListNode<Node> node, Operation operation)
private static void HandleFixedRegisterCopy(LinkedListNode<Node> node, Operation operation)
{
if (operation.SourcesCount == 0)
{
@ -353,7 +353,7 @@ namespace ARMeilleure.CodeGen.X86
Operation storeOp = new Operation(Instruction.Store, null, stackAddr, source);
node.List.AddBefore(node, storeOp);
HandleConstantCopy(node.List.AddBefore(node, storeOp), storeOp);
operation.SetSource(index, stackAddr);
}
@ -390,7 +390,7 @@ namespace ARMeilleure.CodeGen.X86
Operation srcCopyOp = new Operation(Instruction.Copy, argReg, source);
node.List.AddBefore(node, srcCopyOp);
HandleConstantCopy(node.List.AddBefore(node, srcCopyOp), srcCopyOp);
operation.SetSource(index + 1, argReg);
}
@ -405,7 +405,7 @@ namespace ARMeilleure.CodeGen.X86
Operation spillOp = new Operation(Instruction.SpillArg, null, offset, source);
node.List.AddBefore(node, spillOp);
HandleConstantCopy(node.List.AddBefore(node, spillOp), spillOp);
operation.SetSource(index + 1, new Operand(OperandKind.Undefined));
}
@ -444,7 +444,7 @@ namespace ARMeilleure.CodeGen.X86
}
}
private static void AddSameDestSrc1Copy(LinkedListNode<Node> node, Operation operation)
private static void HandleSameDestSrc1Copy(LinkedListNode<Node> node, Operation operation)
{
if (operation.Dest == null || operation.SourcesCount == 0)
{

14
ARMeilleure/Common/BitUtils.cs
View file

@ -67,6 +67,20 @@ namespace ARMeilleure.Common
return output;
}
public static int CountBits(int value)
{
int count = 0;
while (value != 0)
{
value &= ~(value & -value);
count++;
}
return count;
}
public static long FillWithOnes(int bits)
{
return bits == 64 ? -1L : (1L << bits) - 1;

4
ARMeilleure/Instructions/InstEmitMemoryHelper.cs
View file

@ -13,7 +13,7 @@ namespace ARMeilleure.Instructions
{
static class InstEmitMemoryHelper
{
private static bool ForceFallback = false;
private static bool ForceFallback = true;
private enum Extension
{
@ -87,6 +87,8 @@ namespace ARMeilleure.Instructions
case 2: value = context.SignExtend32(value); break;
}
}
context.Copy(GetT(context, rt), value);
}
}

12
ARMeilleure/Instructions/InstEmitSimdCmp.cs
View file

@ -470,9 +470,9 @@ namespace ARMeilleure.Instructions
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Instruction.X86Comissge, n, m);
Operand zf = context.AddIntrinsicInt(Instruction.X86Comisseq, n, m);
Operand nf = context.AddIntrinsicInt(Instruction.X86Comisslt, n, m);
Operand cf = context.AddIntrinsicInt(Instruction.X86Comissge, n, m);
Operand zf = context.AddIntrinsicInt(Instruction.X86Comisseq, n, m);
Operand nf = context.AddIntrinsicInt(Instruction.X86Comisslt, n, m);
context.Copy(GetFlag(PState.VFlag), Const(0));
context.Copy(GetFlag(PState.CFlag), cf);
@ -487,9 +487,9 @@ namespace ARMeilleure.Instructions
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Instruction.X86Comisdge, n, m);
Operand zf = context.AddIntrinsicInt(Instruction.X86Comisdeq, n, m);
Operand nf = context.AddIntrinsicInt(Instruction.X86Comisdlt, n, m);
Operand cf = context.AddIntrinsicInt(Instruction.X86Comisdge, n, m);
Operand zf = context.AddIntrinsicInt(Instruction.X86Comisdeq, n, m);
Operand nf = context.AddIntrinsicInt(Instruction.X86Comisdlt, n, m);
context.Copy(GetFlag(PState.VFlag), Const(0));
context.Copy(GetFlag(PState.CFlag), cf);

40
ARMeilleure/Instructions/SoftFloat.cs
View file

@ -706,8 +706,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static float FPCompareEQ(float value1, float value2, ExecutionContext context)
public static float FPCompareEQ(float value1, float value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -730,8 +732,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static float FPCompareGE(float value1, float value2, ExecutionContext context)
public static float FPCompareGE(float value1, float value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -751,8 +755,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static float FPCompareGT(float value1, float value2, ExecutionContext context)
public static float FPCompareGT(float value1, float value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -772,14 +778,14 @@ namespace ARMeilleure.Instructions
return result;
}
public static float FPCompareLE(float value1, float value2, ExecutionContext context)
public static float FPCompareLE(float value1, float value2)
{
return FPCompareGE(value2, value1, context);
return FPCompareGE(value2, value1);
}
public static float FPCompareLT(float value1, float value2, ExecutionContext context)
public static float FPCompareLT(float value1, float value2)
{
return FPCompareGT(value2, value1, context);
return FPCompareGT(value2, value1);
}
public static float FPDiv(float value1, float value2)
@ -1769,8 +1775,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static double FPCompareEQ(double value1, double value2, ExecutionContext context)
public static double FPCompareEQ(double value1, double value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -1793,8 +1801,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static double FPCompareGE(double value1, double value2, ExecutionContext context)
public static double FPCompareGE(double value1, double value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -1814,8 +1824,10 @@ namespace ARMeilleure.Instructions
return result;
}
public static double FPCompareGT(double value1, double value2, ExecutionContext context)
public static double FPCompareGT(double value1, double value2)
{
ExecutionContext context = NativeInterface.GetContext();
value1 = value1.FPUnpack(out FPType type1, out _, out _, context);
value2 = value2.FPUnpack(out FPType type2, out _, out _, context);
@ -1835,14 +1847,14 @@ namespace ARMeilleure.Instructions
return result;
}
public static double FPCompareLE(double value1, double value2, ExecutionContext context)
public static double FPCompareLE(double value1, double value2)
{
return FPCompareGE(value2, value1, context);
return FPCompareGE(value2, value1);
}
public static double FPCompareLT(double value1, double value2, ExecutionContext context)
public static double FPCompareLT(double value1, double value2)
{
return FPCompareGT(value2, value1, context);
return FPCompareGT(value2, value1);
}
public static double FPDiv(double value1, double value2)