IRInterpreter: Fix issue where we could accidentally optimize out CallReplacement ops.

Core/HLE/ReplaceTables.cpp

@@ -1656,7 +1656,7 @@ void WriteReplaceInstructions(u32 address, u64 hash, int size) {
 	std::vector<int> indexes = GetReplacementFuncIndexes(hash, size);
 	for (int index : indexes) {
 		bool didReplace = false;
-		auto entry = GetReplacementFunc(index);
+		const ReplacementTableEntry *entry = GetReplacementFunc(index);
 		if (entry->flags & REPFLAG_HOOKEXIT) {
 			// When hooking func exit, we search for jr ra, and replace those.
 			for (u32 offset = 0; offset < (u32)size; offset += 4) {

Core/MIPS/IR/IRPassSimplify.cpp

@@ -1175,16 +1175,21 @@ bool PurgeTemps(const IRWriter &in, IRWriter &out, const IROptions &opts) {
 		case IRTEMP_LR_VALUE:
 		case IRTEMP_LR_MASK:
 		case IRTEMP_LR_SHIFT:
-			// Unlike other registers, these don't need to persist between blocks.
-			// So we consider them not read unless proven read.
-			lastWrittenTo[dest] = i;
-			// If this is a copy, we might be able to optimize out the copy.
-			if (inst.op == IROp::Mov) {
-				Check check(dest, i, false);
-				check.srcReg = inst.src1;
-				checks.push_back(check);
+			// Check that it's not a barrier instruction (like CallReplacement). Don't want to even consider optimizing those.
+			if (!(inst.m.flags & IRFLAG_BARRIER)) {
+				// Unlike other registers, these don't need to persist between blocks.
+				// So we consider them not read unless proven read.
+				lastWrittenTo[dest] = i;
+				// If this is a copy, we might be able to optimize out the copy.
+				if (inst.op == IROp::Mov) {
+					Check check(dest, i, false);
+					check.srcReg = inst.src1;
+					checks.push_back(check);
+				} else {
+					checks.push_back(Check(dest, i, false));
+				}
 			} else {
-				checks.push_back(Check(dest, i, false));
+				lastWrittenTo[dest] = i;
 			}
 			break;