Shader: RCP and RSQ computes only the 1st component

src/video_core/shader/shader_interpreter.cpp

@@ -221,13 +221,12 @@ void RunInterpreter(UnitState<Debug>& state) {
             {
                 Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
                 Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+                float24 rcp_res = float24::FromFloat32(1.0f / src1[0].ToFloat32());
                 for (int i = 0; i < 4; ++i) {
                     if (!swizzle.DestComponentEnabled(i))
                         continue;
 
-                    // TODO: Be stable against division by zero!
-                    // TODO: I think this might be wrong... we should only use one component here
-                    dest[i] = float24::FromFloat32(1.0f / src1[i].ToFloat32());
+                    dest[i] = rcp_res;
                 }
                 Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
                 break;
@@ -238,13 +237,12 @@ void RunInterpreter(UnitState<Debug>& state) {
             {
                 Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
                 Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+                float24 rsq_res = float24::FromFloat32(1.0f / sqrt(src1[0].ToFloat32()));
                 for (int i = 0; i < 4; ++i) {
                     if (!swizzle.DestComponentEnabled(i))
                         continue;
 
-                    // TODO: Be stable against division by zero!
-                    // TODO: I think this might be wrong... we should only use one component here
-                    dest[i] = float24::FromFloat32(1.0f / sqrt(src1[i].ToFloat32()));
+                    dest[i] = rsq_res;
                 }
                 Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
                 break;

src/video_core/shader/shader_jit_x64.cpp

@@ -496,9 +496,10 @@ void JitCompiler::Compile_MOV(Instruction instr) {
 void JitCompiler::Compile_RCP(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
-    // TODO(bunnei): RCPPS is a pretty rough approximation, this might cause problems if Pica
+    // TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
     // performs this operation more accurately. This should be checked on hardware.
-    RCPPS(SRC1, R(SRC1));
+    RCPSS(SRC1, R(SRC1));
+    SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
 
     Compile_DestEnable(instr, SRC1);
 }
@@ -506,9 +507,10 @@ void JitCompiler::Compile_RCP(Instruction instr) {
 void JitCompiler::Compile_RSQ(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
-    // TODO(bunnei): RSQRTPS is a pretty rough approximation, this might cause problems if Pica
+    // TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
     // performs this operation more accurately. This should be checked on hardware.
-    RSQRTPS(SRC1, R(SRC1));
+    RSQRTSS(SRC1, R(SRC1));
+    SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
 
     Compile_DestEnable(instr, SRC1);
 }