Jit64: dcbz, use VMOVAPS with YMM registers on AVX CPUs

LLVM does this, so let's do it as well.

Source/Core/Common/x64Emitter.cpp

@@ -1858,19 +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,
+void CheckAVXSupport()
-                          int W, int extrabytes)
 {
   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)
+  CheckAVXSupport();
-    PanicAlertFmt("Trying to use AVX on a system that doesn't support it. Bad programmer.");
   WriteVEXOp4(opPrefix, op, regOp1, regOp2, arg, regOp3, W);
 }
 
@@ -3030,6 +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/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"
@@ -469,9 +470,18 @@ 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));
+    if (cpu_info.bAVX)
-    MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), XMM0);
+    {
-    MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 16), XMM0);
+      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();