Merge pull request #13929 from SuperSamus/dcbz-jit-improvements

Jit: Small `dcbz` improvements

Source/Core/Common/x64Emitter.cpp

@@ -250,7 +250,7 @@ void OpArg::WriteVEX(XEmitter* emit, X64Reg regOp1, X64Reg regOp2, int L, int pp
   int X = !(indexReg & 8);
   int B = !(offsetOrBaseReg & 8);
 
-  int vvvv = (regOp2 == X64Reg::INVALID_REG) ? 0xf : (regOp2 ^ 0xf);
+  u8 vvvv = (regOp2 == X64Reg::INVALID_REG) ? 0xf : (regOp2 ^ 0xf);
 
   // do we need any VEX fields that only appear in the three-byte form?
   if (X == 1 && B == 1 && W == 0 && mmmmm == 1)
@@ -343,7 +343,7 @@ void OpArg::WriteRest(XEmitter* emit, int extraBytes, X64Reg _operandReg,
   if (SIB)
     oreg = 4;
 
-  emit->WriteModRM(mod, _operandReg & 7, oreg & 7);
+  emit->WriteModRM(mod, _operandReg, oreg);
 
   if (SIB)
   {
@@ -1844,8 +1844,9 @@ void XEmitter::WriteVEXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, con
 {
   int mmmmm = GetVEXmmmmm(op);
   int pp = GetVEXpp(opPrefix);
-  // FIXME: we currently don't support 256-bit instructions, and "size" is not the vector size here
-  arg.WriteVEX(this, regOp1, regOp2, 0, pp, mmmmm, W);
+  // Note that mixing an XMM register with a YMM register is invalid, which isn't checked here.
+  int L = (regOp1 != INVALID_REG && regOp1 & 0x100) || (regOp2 != INVALID_REG && regOp2 & 0x100);
+  arg.WriteVEX(this, regOp1, regOp2, L, pp, mmmmm, W);
   Write8(op & 0xFF);
   arg.WriteRest(this, extrabytes, regOp1);
 }
@@ -1857,19 +1858,23 @@ void XEmitter::WriteVEXOp4(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, co
   Write8((u8)regOp3 << 4);
 }
 
-void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
-                          int W, int extrabytes)
+void CheckAVXSupport()
 {
   if (!cpu_info.bAVX)
     PanicAlertFmt("Trying to use AVX on a system that doesn't support it. Bad programmer.");
+}
+
+void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
+                          int W, int extrabytes)
+{
+  CheckAVXSupport();
   WriteVEXOp(opPrefix, op, regOp1, regOp2, arg, W, extrabytes);
 }
 
 void XEmitter::WriteAVXOp4(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
                            X64Reg regOp3, int W)
 {
-  if (!cpu_info.bAVX)
-    PanicAlertFmt("Trying to use AVX on a system that doesn't support it. Bad programmer.");
+  CheckAVXSupport();
   WriteVEXOp4(opPrefix, op, regOp1, regOp2, arg, regOp3, W);
 }
 
@@ -3029,6 +3034,19 @@ void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg)
   WriteAVXOp(0x66, 0xEF, regOp1, regOp2, arg);
 }
 
+void XEmitter::VMOVAPS(const OpArg& arg, X64Reg regOp)
+{
+  WriteAVXOp(0x00, 0x29, X64Reg::INVALID_REG, regOp, arg);
+}
+
+void XEmitter::VZEROUPPER()
+{
+  CheckAVXSupport();
+  Write8(0xC5);
+  Write8(0xF8);
+  Write8(0x77);
+}
+
 void XEmitter::VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg)
 {
   WriteFMA3Op(0x98, regOp1, regOp2, arg);

Source/Core/Common/x64Emitter.h

@@ -876,6 +876,10 @@ public:
   void VPOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
   void VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
 
+  void VMOVAPS(const OpArg& arg, X64Reg regOp);
+
+  void VZEROUPPER();
+
   // FMA3
   void VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
   void VFMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);

Source/Core/Common/x64Reg.h

@@ -72,7 +72,8 @@ enum X64Reg
   XMM14,
   XMM15,
 
-  YMM0 = 0,
+  // Use the bit 0x100 to distinguish XMM and YMM registers.
+  YMM0 = 0x100,
   YMM1,
   YMM2,
   YMM3,

Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp

@@ -101,7 +101,7 @@ constexpr std::array<Jit64OpTemplate, 13> s_table4{{
     {592, &Jit64::ps_mergeXX},  // ps_merge10
     {624, &Jit64::ps_mergeXX},  // ps_merge11
 
-    {1014, &Jit64::FallBackToInterpreter},  // dcbz_l
+    {1014, &Jit64::dcbz},  // dcbz_l
 }};
 
 constexpr std::array<Jit64OpTemplate, 17> s_table4_2{{

Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -8,6 +8,7 @@
 
 #include "Common/Assert.h"
 #include "Common/BitSet.h"
+#include "Common/CPUDetect.h"
 #include "Common/CommonTypes.h"
 #include "Common/MsgHandler.h"
 #include "Common/x64ABI.h"
@@ -473,9 +474,18 @@ void Jit64::dcbz(UGeckoInstruction inst)
     FixupBranch slow = J_CC(CC_Z, Jump::Near);
 
     // Fast path: compute full address, then zero out 32 bytes of memory.
-    XORPS(XMM0, R(XMM0));
-    MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), XMM0);
-    MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 16), XMM0);
+    if (cpu_info.bAVX)
+    {
+      VXORPS(XMM0, XMM0, R(XMM0));
+      VMOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), YMM0);
+      VZEROUPPER();
+    }
+    else
+    {
+      XORPS(XMM0, R(XMM0));
+      MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 16), XMM0);
+      MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), XMM0);
+    }
 
     // Slow path: call the general-case code.
     SwitchToFarCode();

Source/Core/Core/PowerPC/JitArm64/JitArm64_Tables.cpp

@@ -101,7 +101,7 @@ constexpr std::array<JitArm64OpTemplate, 13> s_table4{{
     {592, &JitArm64::ps_mergeXX},  // ps_merge10
     {624, &JitArm64::ps_mergeXX},  // ps_merge11
 
-    {1014, &JitArm64::FallBackToInterpreter},  // dcbz_l
+    {1014, &JitArm64::dcbz},  // dcbz_l
 }};
 
 constexpr std::array<JitArm64OpTemplate, 17> s_table4_2{{