microVU: Move VU0 micro flag instance setup to program start

They're only used in micro mode, so no point updating them in a cop2
chain.

pcsx2/VU0micro.cpp

@@ -34,6 +34,21 @@ void vu0ResetRegs()
 	vif0Regs.stat.VEW = false;
 }
 
+static __fi u32 vu0DenormalizeMicroStatus(u32 nstatus)
+{
+	// from mVUallocSFLAGd()
+	return ((nstatus >> 3) & 0x18u) | ((nstatus >> 11) & 0x1800u) | ((nstatus >> 14) & 0x3cf0000u);
+}
+
+static __fi void vu0SetMicroFlags(u32* flags, u32 value)
+{
+#ifdef _M_X86
+	_mm_store_si128(reinterpret_cast<__m128i*>(flags), _mm_set1_epi32(value));
+#else
+	flags[0] = flags[1] = flags[2] = flags[3] = value;
+#endif
+}
+
 void __fastcall vu0ExecMicro(u32 addr) {
 	VUM_LOG("vu0ExecMicro %x", addr);
 
@@ -43,9 +58,20 @@ void __fastcall vu0ExecMicro(u32 addr) {
 	}
 
 	// Need to copy the clip flag back to the interpreter in case COP2 has edited it
-	VU0.clipflag = VU0.VI[REG_CLIP_FLAG].UL;
+	const u32 CLIP = VU0.VI[REG_CLIP_FLAG].UL;
-	VU0.macflag = VU0.VI[REG_MAC_FLAG].UL;
+	const u32 MAC = VU0.VI[REG_MAC_FLAG].UL;
-	VU0.statusflag = VU0.VI[REG_STATUS_FLAG].UL;
+	const u32 STATUS = VU0.VI[REG_STATUS_FLAG].UL;
+	VU0.clipflag = CLIP;
+	VU0.macflag = MAC;
+	VU0.statusflag = STATUS;
+
+	// Copy flags to micro instances, since they may be out of sync if COP2 has run.
+	// We do this at program start time, because COP2 can't execute until the program has completed,
+	// but long-running program may be interrupted so we can't do it at dispatch time.
+	vu0SetMicroFlags(VU0.micro_clipflags, CLIP);
+	vu0SetMicroFlags(VU0.micro_macflags, MAC);
+	vu0SetMicroFlags(VU0.micro_statusflags, vu0DenormalizeMicroStatus(STATUS));
+
 	VU0.VI[REG_VPU_STAT].UL &= ~0xFF;
 	VU0.VI[REG_VPU_STAT].UL |=  0x01;
 	VU0.cycle = cpuRegs.cycle;

pcsx2/x86/microVU_Alloc.inl

@@ -110,16 +110,6 @@ __fi void mVUallocCFLAGb(mV, const x32& reg, int fInstance)
 {
 	if (fInstance < 4) xMOV(ptr32[&mVU.clipFlag[fInstance]], reg);         // microVU
 	else               xMOV(ptr32[&mVU.regs().VI[REG_CLIP_FLAG].UL], reg); // macroVU
-
-	// On COP2 modifying the CLIP flag we need to update the microVU version for when it's restored on new program
-	if (fInstance == 0xff)
-	{
-		int t0reg = _allocTempXMMreg(XMMT_INT, -1);
-		xMOVDZX(xRegisterSSE(t0reg), reg);
-		xPSHUF.D(xRegisterSSE(t0reg), xRegisterSSE(t0reg), 0);
-		xMOVDQA(ptr128[&mVU.regs().micro_clipflags], xRegisterSSE(t0reg));
-		_freeXMMreg(t0reg);
-	}
 }
 
 //------------------------------------------------------------------

pcsx2/x86/microVU_Macro.inl

@@ -102,16 +102,6 @@ void endMacroOp(int mode)
 
 	if (mode & 0x10)
 	{
-		if (!CHECK_VU_FLAGHACK || (g_pCurInstInfo->info & EEINST_COP2_STATUS_FLAG)) // Status/Mac Flags were Updated
-		{
-			// update micro_statusflags
-			const int t0reg = _allocTempXMMreg(XMMT_INT, -1);
-			xMOVDZX(xRegisterSSE(t0reg), gprF0);
-			xSHUF.PS(xRegisterSSE(t0reg), xRegisterSSE(t0reg), 0);
-			xMOVAPS(ptr128[&microVU0.regs().micro_statusflags], xRegisterSSE(t0reg));
-			_freeXMMreg(t0reg);
-		}
-
 		if (!CHECK_VU_FLAGHACK || g_pCurInstInfo->info & EEINST_COP2_NORMALIZE_STATUS_FLAG)
 		{
 			// Normalize
@@ -124,15 +114,6 @@ void endMacroOp(int mode)
 			// this is fine, because we'll normalize them again before this reg is accessed
 			xMOV(ptr32[&vuRegs->VI[REG_STATUS_FLAG].UL], gprF0);
 		}
-
-		if (!CHECK_VU_FLAGHACK || (g_pCurInstInfo->info & EEINST_COP2_MAC_FLAG))
-		{
-			const int t0reg = _allocTempXMMreg(XMMT_INT, -1);
-			xMOVDZX(xRegisterSSE(t0reg), ptr32[&vu0Regs.VI[REG_MAC_FLAG].UL]);
-			xSHUF.PS(xRegisterSSE(t0reg), xRegisterSSE(t0reg), 0);
-			xMOVAPS(ptr128[&microVU0.regs().micro_macflags], xRegisterSSE(t0reg));
-			_freeXMMreg(t0reg);
-		}
 	}
 
 	microVU0.cop2 = 0;