Misc UE4 fixes (#1821)

* Add ExecLo case to S_SAVEEXEC_B64

Seen in CUSA38209

* S_BCNT1_I32_B32

Turtle said our implementation of S_BCNT1_I32_B64 was meant to be for S_BCNT1_I32_B32, so renaming the opcode is the fix.

src/shader_recompiler/frontend/translate/scalar_alu.cpp

@@ -98,8 +98,8 @@ void Translator::EmitScalarAlu(const GcnInst& inst) {
             break;
         case Opcode::S_BREV_B32:
             return S_BREV_B32(inst);
-        case Opcode::S_BCNT1_I32_B64:
-            return S_BCNT1_I32_B64(inst);
+        case Opcode::S_BCNT1_I32_B32:
+            return S_BCNT1_I32_B32(inst);
         case Opcode::S_FF1_I32_B32:
             return S_FF1_I32_B32(inst);
         case Opcode::S_AND_SAVEEXEC_B64:
@@ -579,7 +579,7 @@ void Translator::S_BREV_B32(const GcnInst& inst) {
     SetDst(inst.dst[0], ir.BitReverse(GetSrc(inst.src[0])));
 }
 
-void Translator::S_BCNT1_I32_B64(const GcnInst& inst) {
+void Translator::S_BCNT1_I32_B32(const GcnInst& inst) {
     const IR::U32 result = ir.BitCount(GetSrc(inst.src[0]));
     SetDst(inst.dst[0], result);
     ir.SetScc(ir.INotEqual(result, ir.Imm32(0)));
@@ -602,6 +602,8 @@ void Translator::S_SAVEEXEC_B64(NegateMode negate, bool is_or, const GcnInst& in
             return ir.GetVcc();
         case OperandField::ScalarGPR:
             return ir.GetThreadBitScalarReg(IR::ScalarReg(inst.src[0].code));
+        case OperandField::ExecLo:
+            return ir.GetExec();
         default:
             UNREACHABLE();
         }

src/shader_recompiler/frontend/translate/translate.h

@@ -110,7 +110,7 @@ class Translator {
     void S_MOV_B64(const GcnInst& inst);
     void S_NOT_B64(const GcnInst& inst);
     void S_BREV_B32(const GcnInst& inst);
-    void S_BCNT1_I32_B64(const GcnInst& inst);
+    void S_BCNT1_I32_B32(const GcnInst& inst);
     void S_FF1_I32_B32(const GcnInst& inst);
     void S_GETPC_B64(u32 pc, const GcnInst& inst);
     void S_SAVEEXEC_B64(NegateMode negate, bool is_or, const GcnInst& inst);