diff --git a/Core/HLE/sceKernel.cpp b/Core/HLE/sceKernel.cpp
index d4da98bf42..e0427b9ec8 100644
--- a/Core/HLE/sceKernel.cpp
+++ b/Core/HLE/sceKernel.cpp
@@ -322,7 +322,7 @@ void sceKernelExitGame()
 	__KernelSwitchOffThread("game exited");
 	Core_Stop();
 
-	g_OSD.Show(OSDType::MESSAGE_INFO, "sceKernelExitGame()");
+	g_OSD.Show(OSDType::MESSAGE_INFO, "sceKernelExitGame()", 0.0f, "kernelexit");
 }
 
 void sceKernelExitGameWithStatus()
diff --git a/Core/MIPS/IR/IRCompFPU.cpp b/Core/MIPS/IR/IRCompFPU.cpp
index e83f3a709f..ca603199ea 100644
--- a/Core/MIPS/IR/IRCompFPU.cpp
+++ b/Core/MIPS/IR/IRCompFPU.cpp
@@ -223,6 +223,7 @@ void IRFrontend::Comp_mxc1(MIPSOpcode op) {
 			// Set rounding mode
 			RestoreRoundingMode();
 			ir.Write(IROp::FpCtrlFromReg, 0, rt);
+			// TODO: Do the UpdateRoundingMode check at runtime?
 			UpdateRoundingMode();
 			ApplyRoundingMode();
 		} else {
diff --git a/Core/MIPS/IR/IRFrontend.cpp b/Core/MIPS/IR/IRFrontend.cpp
index 4af6d9a753..f299ba9cae 100644
--- a/Core/MIPS/IR/IRFrontend.cpp
+++ b/Core/MIPS/IR/IRFrontend.cpp
@@ -133,11 +133,9 @@ bool IRFrontend::CheckRounding(u32 blockAddress) {
 		js.startDefaultPrefix = false;
 		cleanSlate = true;
 	}
-
 	return cleanSlate;
 }
 
-
 void IRFrontend::Comp_ReplacementFunc(MIPSOpcode op) {
 	int index = op.encoding & MIPS_EMUHACK_VALUE_MASK;
 
diff --git a/Core/MIPS/IR/IRInterpreter.cpp b/Core/MIPS/IR/IRInterpreter.cpp
index 4d46983c97..bbabadefa1 100644
--- a/Core/MIPS/IR/IRInterpreter.cpp
+++ b/Core/MIPS/IR/IRInterpreter.cpp
@@ -89,6 +89,37 @@ u32 IRRunMemCheck(u32 pc, u32 addr) {
 	return coreState != CORE_RUNNING ? 1 : 0;
 }
 
+void IRApplyRounding(MIPSState *mips) {
+	u32 fcr1Bits = mips->fcr31 & 0x01000003;
+	// If these are 0, we just leave things as they are.
+	if (fcr1Bits) {
+#if PPSSPP_ARCH(SSE2) && 0
+		u32 csr = _mm_getcsr() & ~0x6000;
+		// Translate the rounding mode bits to X86, the same way as in Asm.cpp.
+		int rnd = fcr1Bits & 3;
+		if (rnd & 1) {
+			rnd ^= 2;
+		}
+		csr |= rnd << 13;
+
+		if (fcr1Bits & 0x01000000) {
+			// Flush to zero
+			csr |= 0x8000;
+		}
+		_mm_setcsr(csr);
+#endif
+	}
+}
+
+void IRRestoreRounding() {
+#if PPSSPP_ARCH(SSE2)
+	// We should avoid this if we didn't apply rounding in the first place.
+	u32 csr = _mm_getcsr();
+	csr &= ~(7 << 13);
+	_mm_setcsr(csr);
+#endif
+}
+
 // We cannot use NEON on ARM32 here until we make it a hard dependency. We can, however, on ARM64.
 u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 	while (true) {
@@ -565,9 +596,11 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			}
 			break;
 
-		// Not quickly implementable on all platforms, unfortunately.
 		case IROp::Vec4Dot:
 		{
+			// Not quickly implementable on all platforms, unfortunately.
+			// Though, this is still pretty fast compared to one split into multiple IR instructions.
+			// This might be good though: https://stackoverflow.com/a/17004629
 			float dot = mips->f[inst->src1] * mips->f[inst->src2];
 			for (int i = 1; i < 4; i++)
 				dot += mips->f[inst->src1 + i] * mips->f[inst->src2 + i];
@@ -826,9 +859,9 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			mips->f[inst->dest] = vfpu_clamp(mips->f[inst->src1], -1.0f, 1.0f);
 			break;
 
-		// Bitwise trickery
 		case IROp::FSign:
 		{
+			// Bitwise trickery
 			u32 val;
 			memcpy(&val, &mips->f[inst->src1], sizeof(u32));
 			if (val == 0 || val == 0x80000000)
@@ -1097,9 +1130,11 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			break;
 
 		case IROp::ApplyRoundingMode:
+			IRApplyRounding(mips);
 			// TODO: Implement
 			break;
 		case IROp::RestoreRoundingMode:
+			IRRestoreRounding();
 			// TODO: Implement
 			break;
 		case IROp::UpdateRoundingMode:
diff --git a/Core/MIPS/IR/IRInterpreter.h b/Core/MIPS/IR/IRInterpreter.h
index 6f02ec9482..0939e982d0 100644
--- a/Core/MIPS/IR/IRInterpreter.h
+++ b/Core/MIPS/IR/IRInterpreter.h
@@ -11,6 +11,9 @@ u32 IRRunBreakpoint(u32 pc);
 u32 IRRunMemCheck(u32 pc, u32 addr);
 u32 IRInterpret(MIPSState *ms, const IRInst *inst);
 
+void IRApplyRounding();
+void IRRestoreRounding();
+
 template <uint32_t alignment>
 u32 RunValidateAddress(u32 pc, u32 addr, u32 isWrite) {
 	const auto toss = [&](MemoryExceptionType t) {