diff --git a/src/common/bit_set.h b/src/common/bit_set.h
index 9c2e6b28c0..749de4df01 100644
--- a/src/common/bit_set.h
+++ b/src/common/bit_set.h
@@ -120,20 +120,15 @@ public:
     // A STL-like iterator is required to be able to use range-based for loops.
     class Iterator {
     public:
-        Iterator(const Iterator& other) : m_val(other.m_val), m_bit(other.m_bit) {}
-        Iterator(IntTy val) : m_val(val), m_bit(0) {}
-        Iterator& operator=(Iterator other) {
-            new (this) Iterator(other);
-            return *this;
-        }
+        Iterator(const Iterator& other) : m_val(other.m_val) {}
+        Iterator(IntTy val) : m_val(val) {}
         int operator*() {
-            return m_bit + ComputeLsb();
+            // This will never be called when m_val == 0, because that would be the end() iterator
+            return LeastSignificantSetBit(m_val);
         }
         Iterator& operator++() {
-            int lsb = ComputeLsb();
-            m_val >>= lsb + 1;
-            m_bit += lsb + 1;
-            m_has_lsb = false;
+            // Unset least significant set bit
+            m_val &= m_val - IntTy(1);
             return *this;
         }
         Iterator operator++(int _) {
@@ -149,17 +144,7 @@ public:
         }
 
     private:
-        int ComputeLsb() {
-            if (!m_has_lsb) {
-                m_lsb = LeastSignificantSetBit(m_val);
-                m_has_lsb = true;
-            }
-            return m_lsb;
-        }
         IntTy m_val;
-        int m_bit;
-        int m_lsb = -1;
-        bool m_has_lsb = false;
     };
 
     BitSet() : m_val(0) {}
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index 2857d28297..2d0ffe8216 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -77,13 +77,18 @@ void UnitState::LoadInput(const ShaderRegs& config, const AttributeBuffer& input
     }
 }
 
-void UnitState::WriteOutput(const ShaderRegs& config, AttributeBuffer& output) {
-    unsigned int output_i = 0;
-    for (unsigned int reg : Common::BitSet<u32>(config.output_mask)) {
-        output.attr[output_i++] = registers.output[reg];
+static void CopyRegistersToOutput(const Math::Vec4<float24>* regs, u32 mask,
+                                  AttributeBuffer& buffer) {
+    int output_i = 0;
+    for (int reg : Common::BitSet<u32>(mask)) {
+        buffer.attr[output_i++] = regs[reg];
     }
 }
 
+void UnitState::WriteOutput(const ShaderRegs& config, AttributeBuffer& output) {
+    CopyRegistersToOutput(registers.output, config.output_mask, output);
+}
+
 UnitState::UnitState(GSEmitter* emitter) : emitter_ptr(emitter) {}
 
 GSEmitter::GSEmitter() {
@@ -94,19 +99,16 @@ GSEmitter::~GSEmitter() {
     delete handlers;
 }
 
-void GSEmitter::Emit(Math::Vec4<float24> (&vertex)[16]) {
+void GSEmitter::Emit(Math::Vec4<float24> (&output_regs)[16]) {
     ASSERT(vertex_id < 3);
-    std::copy(std::begin(vertex), std::end(vertex), buffer[vertex_id].begin());
+    // TODO: This should be merged with UnitState::WriteOutput somehow
+    CopyRegistersToOutput(output_regs, output_mask, buffer[vertex_id]);
+
     if (prim_emit) {
         if (winding)
             handlers->winding_setter();
         for (size_t i = 0; i < buffer.size(); ++i) {
-            AttributeBuffer output;
-            unsigned int output_i = 0;
-            for (unsigned int reg : Common::BitSet<u32>(output_mask)) {
-                output.attr[output_i++] = buffer[i][reg];
-            }
-            handlers->vertex_handler(output);
+            handlers->vertex_handler(buffer[i]);
         }
     }
 }
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index a3789da012..8740a16189 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -72,7 +72,7 @@ static_assert(sizeof(OutputVertex) == 24 * sizeof(float), "OutputVertex has inva
  * This structure contains state information for primitive emitting in geometry shader.
  */
 struct GSEmitter {
-    std::array<std::array<Math::Vec4<float24>, 16>, 3> buffer;
+    std::array<AttributeBuffer, 3> buffer;
     u8 vertex_id;
     bool prim_emit;
     bool winding;
@@ -87,7 +87,7 @@ struct GSEmitter {
 
     GSEmitter();
     ~GSEmitter();
-    void Emit(Math::Vec4<float24> (&vertex)[16]);
+    void Emit(Math::Vec4<float24> (&output_regs)[16]);
 };
 static_assert(std::is_standard_layout<GSEmitter>::value, "GSEmitter is not standard layout type");