diff --git a/src/video_core/renderer_opengl/gl_arb_decompiler.cpp b/src/video_core/renderer_opengl/gl_arb_decompiler.cpp
index a1f5a1f8c3..53cfe73ba6 100644
--- a/src/video_core/renderer_opengl/gl_arb_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_arb_decompiler.cpp
@@ -1178,11 +1178,7 @@ void ARBDecompiler::VisitAST(const ASTNode& node) {
         if (ast_return->kills) {
             AddLine("KIL TR;");
         } else {
-            if (context_func->IsMain()) {
-                Exit();
-            } else {
-                AddLine("RET;");
-            }
+            Exit();
         }
         if (!is_true) {
             AddLine("ENDIF;");
@@ -1487,7 +1483,7 @@ std::string ARBDecompiler::GlobalMemoryPointer(const GmemNode& gmem) {
 }
 
 void ARBDecompiler::Exit() {
-    if (stage != ShaderType::Fragment) {
+    if (!context_func->IsMain() || stage != ShaderType::Fragment) {
         AddLine("RET;");
         return;
     }
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index 7d9d627a42..5404a388ea 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -435,6 +435,25 @@ public:
         DeclareCustomVariables();
         DeclarePhysicalAttributeReader();
 
+        const auto& subfunctions = ir.GetSubFunctions();
+        auto it = subfunctions.rbegin();
+        while (it != subfunctions.rend()) {
+            context_func = *it;
+            code.AddLine("void func_{}() {{", context_func->GetId());
+            ++code.scope;
+
+            if (context_func->IsDecompiled()) {
+                DecompileAST();
+            } else {
+                DecompileBranchMode();
+            }
+
+            --code.scope;
+            code.AddLine("}}");
+
+            it++;
+        }
+
         context_func = ir.GetMainFunction();
 
         code.AddLine("void main() {{");
@@ -1133,6 +1152,11 @@ private:
             return {};
         }
 
+        if (const auto func_call = std::get_if<FunctionCallNode>(&*node)) {
+            code.AddLine("func_{}();", func_call->GetFuncId());
+            return {};
+        }
+
         if (const auto comment = std::get_if<CommentNode>(&*node)) {
             code.AddLine("// " + comment->GetText());
             return {};
@@ -2269,7 +2293,9 @@ private:
     }
 
     Expression Exit(Operation operation) {
-        PreExit();
+        if (context_func->IsMain()) {
+            PreExit();
+        }
         code.AddLine("return;");
         return {};
     }
diff --git a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
index 4ead6c5f8b..7e8a228e3d 100644
--- a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
+++ b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
@@ -2177,7 +2177,9 @@ private:
     }
 
     Expression Exit(Operation operation) {
-        PreExit();
+        if (context_func->IsMain()) {
+            PreExit();
+        }
         inside_branch = true;
         if (conditional_branch_set) {
             OpReturn();
@@ -3066,7 +3068,9 @@ public:
             if (ast.kills) {
                 decomp.OpKill();
             } else {
-                decomp.PreExit();
+                if (decomp.context_func->IsMain()) {
+                    decomp.PreExit();
+                }
                 decomp.OpReturn();
             }
             decomp.AddLabel(decomp.OpLabel());
diff --git a/src/video_core/shader/control_flow.h b/src/video_core/shader/control_flow.h
index 33ecb7cf39..5ef2251b95 100644
--- a/src/video_core/shader/control_flow.h
+++ b/src/video_core/shader/control_flow.h
@@ -5,6 +5,7 @@
 #pragma once
 
 #include <list>
+#include <map>
 #include <optional>
 #include <set>
 #include <variant>
@@ -112,12 +113,10 @@ struct ShaderFunction {
 
 struct ShaderProgram {
     ShaderFunction main;
-    std::unordered_map<u32, ShaderFunction> subfunctions;
+    std::map<u32, ShaderFunction> subfunctions;
 };
 
 std::unique_ptr<ShaderProgram> ScanFlow(const ProgramCode& program_code, u32 start_address,
-                                                const CompilerSettings& settings,
-                                                Registry& registry);
-
+                                        const CompilerSettings& settings, Registry& registry);
 
 } // namespace VideoCommon::Shader
diff --git a/src/video_core/shader/decode.cpp b/src/video_core/shader/decode.cpp
index eaa8b46bb0..4667c37317 100644
--- a/src/video_core/shader/decode.cpp
+++ b/src/video_core/shader/decode.cpp
@@ -216,10 +216,22 @@ void ShaderIR::Decode() {
 
     decompiled = false;
     auto info = ScanFlow(program_code, main_offset, settings, registry);
+    u32 id_start = 1;
+    for (auto& pair : info->subfunctions) {
+        func_map.emplace(pair.first, id_start);
+        id_start++;
+    }
     coverage_begin = info->main.start;
     coverage_end = 0;
     decode_function(info->main);
     main_function = gen_function(info->main, 0);
+    subfunctions.resize(info->subfunctions.size());
+    for (auto& pair : info->subfunctions) {
+        auto& func_info = pair.second;
+        decode_function(func_info);
+        u32 id = func_map[pair.first];
+        subfunctions[id - 1] = gen_function(func_info, id);
+    }
 }
 
 NodeBlock ShaderIR::DecodeRange(u32 begin, u32 end) {
diff --git a/src/video_core/shader/decode/other.cpp b/src/video_core/shader/decode/other.cpp
index 5f88537bc4..2bc596512a 100644
--- a/src/video_core/shader/decode/other.cpp
+++ b/src/video_core/shader/decode/other.cpp
@@ -33,6 +33,7 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         // With the previous preconditions, this instruction is a no-operation.
         break;
     }
+    case OpCode::Id::RET:
     case OpCode::Id::EXIT: {
         const ConditionCode cc = instr.flow_condition_code;
         UNIMPLEMENTED_IF_MSG(cc != ConditionCode::T, "EXIT condition code used: {}", cc);
@@ -312,6 +313,16 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         LOG_DEBUG(HW_GPU, "DEPBAR instruction is stubbed");
         break;
     }
+    case OpCode::Id::CAL: {
+        const u32 target = pc + instr.bra.GetBranchTarget();
+        const auto it = func_map.find(target);
+        if (it == func_map.end()) {
+            UNREACHABLE();
+            break;
+        }
+        bb.push_back(FunctionCall(it->second));
+        break;
+    }
     default:
         UNIMPLEMENTED_MSG("Unhandled instruction: {}", opcode->get().GetName());
     }
diff --git a/src/video_core/shader/node.h b/src/video_core/shader/node.h
index b54d33763d..a58e7c65e4 100644
--- a/src/video_core/shader/node.h
+++ b/src/video_core/shader/node.h
@@ -267,10 +267,11 @@ class PatchNode;
 class SmemNode;
 class GmemNode;
 class CommentNode;
+class FunctionCallNode;
 
 using NodeData = std::variant<OperationNode, ConditionalNode, GprNode, CustomVarNode, ImmediateNode,
                               InternalFlagNode, PredicateNode, AbufNode, PatchNode, CbufNode,
-                              LmemNode, SmemNode, GmemNode, CommentNode>;
+                              LmemNode, SmemNode, GmemNode, FunctionCallNode, CommentNode>;
 using Node = std::shared_ptr<NodeData>;
 using Node4 = std::array<Node, 4>;
 using NodeBlock = std::vector<Node>;
@@ -494,6 +495,18 @@ private:
     std::vector<Node> code; ///< Code to execute
 };
 
+class FunctionCallNode final : public AmendNode {
+public:
+    explicit FunctionCallNode(u32 func_id_) : func_id{func_id_} {}
+
+    [[nodiscard]] u32 GetFuncId() const {
+        return func_id;
+    }
+
+private:
+    u32 func_id; ///< Id of the function to call
+};
+
 /// A general purpose register
 class GprNode final {
 public:
diff --git a/src/video_core/shader/node_helper.cpp b/src/video_core/shader/node_helper.cpp
index 6a5b6940d1..cef9c26bc3 100644
--- a/src/video_core/shader/node_helper.cpp
+++ b/src/video_core/shader/node_helper.cpp
@@ -19,6 +19,11 @@ Node Comment(std::string text) {
     return MakeNode<CommentNode>(std::move(text));
 }
 
+/// Creates a function call
+Node FunctionCall(u32 func_id) {
+    return MakeNode<FunctionCallNode>(func_id);
+}
+
 Node Immediate(u32 value) {
     return MakeNode<ImmediateNode>(value);
 }
diff --git a/src/video_core/shader/node_helper.h b/src/video_core/shader/node_helper.h
index 1e0886185d..3f882cd25d 100644
--- a/src/video_core/shader/node_helper.h
+++ b/src/video_core/shader/node_helper.h
@@ -27,6 +27,9 @@ Node Conditional(Node condition, std::vector<Node> code);
 /// Creates a commentary node
 Node Comment(std::string text);
 
+/// Creates a function call
+Node FunctionCall(u32 func_id);
+
 /// Creates an u32 immediate
 Node Immediate(u32 value);
 
diff --git a/src/video_core/shader/shader_ir.h b/src/video_core/shader/shader_ir.h
index b7a33d0ae8..56983c9479 100644
--- a/src/video_core/shader/shader_ir.h
+++ b/src/video_core/shader/shader_ir.h
@@ -69,7 +69,7 @@ public:
     explicit ShaderFunctionIR(std::map<u32, NodeBlock>&& basic_blocks_, bool disable_flow_stack_,
                               u32 id_, u32 coverage_begin_, u32 coverage_end_)
         : basic_blocks{std::move(basic_blocks_)}, decompiled{false},
-          disable_flow_stack{disable_flow_stack}, id{id_}, coverage_begin{coverage_begin_},
+          disable_flow_stack{disable_flow_stack_}, id{id_}, coverage_begin{coverage_begin_},
           coverage_end{coverage_end_} {}
     explicit ShaderFunctionIR(ASTManager&& program_manager_, u32 id_, u32 coverage_begin_,
                               u32 coverage_end_)
@@ -80,11 +80,11 @@ public:
         return basic_blocks;
     }
 
-    bool IsFlowStackDisabled() const {
+    [[nodiscard]] bool IsFlowStackDisabled() const {
         return disable_flow_stack;
     }
 
-    bool IsDecompiled() const {
+    [[nodiscard]] bool IsDecompiled() const {
         return decompiled;
     }
 
@@ -100,11 +100,11 @@ public:
         return program_manager.GetVariables();
     }
 
-    bool IsMain() const {
+    [[nodiscard]] bool IsMain() const {
         return id == 0;
     }
 
-    u32 GetId() const {
+    [[nodiscard]] u32 GetId() const {
         return id;
     }
 
@@ -226,6 +226,10 @@ public:
         return main_function;
     }
 
+    const std::vector<std::shared_ptr<ShaderFunctionIR>>& GetSubFunctions() const {
+        return subfunctions;
+    }
+
 private:
     friend class ASTDecoder;
 
@@ -491,6 +495,7 @@ private:
 
     std::shared_ptr<ShaderFunctionIR> main_function;
     std::vector<std::shared_ptr<ShaderFunctionIR>> subfunctions;
+    std::unordered_map<u32, u32> func_map;
 
     std::set<u32> used_registers;
     std::set<Tegra::Shader::Pred> used_predicates;