diff --git a/ps2xRecomp/src/lib/function_emitter.cpp b/ps2xRecomp/src/lib/function_emitter.cpp
index 328c0e01a..f4a80b4db 100644
--- a/ps2xRecomp/src/lib/function_emitter.cpp
+++ b/ps2xRecomp/src/lib/function_emitter.cpp
@@ -40,20 +40,6 @@ namespace ps2recomp
         std::stringstream ss;
         cg.m_currentFunctionName = function.name;
 
-        if (useHeaders)
-        {
-            ss << "#include <stdexcept>\n";
-            ss << "#include \"ps2_runtime_macros.h\"\n";
-            ss << "#include \"ps2_runtime.h\"\n";
-            ss << "#include \"ps2_recompiled_functions.h\"\n";
-            ss << "#include \"ps2_recompiled_stubs.h\"\n\n";
-            ss << "#include \"ps2_syscalls.h\"\n";
-            ss << "#include \"ps2_stubs.h\"\n\n";
-            ss << "#ifdef PS2_FUNCTION_LOG_TRACKER\n";
-            ss << "#include \"ps2_log.h\"\n";
-            ss << "#endif\n\n";
-        }
-
         CodeGenerator::AnalysisResult analysisResult = cg.collectInternalBranchTargets(function, instructions);
         std::vector<uint32_t> resumeTargets(analysisResult.resumeEntryPoints.begin(),
                                             analysisResult.resumeEntryPoints.end());
@@ -234,6 +220,50 @@ namespace ps2recomp
         }
 
         ss << "}\n";
-        return ss.str();
+
+        const std::string body = ss.str();
+        if (!useHeaders)
+        {
+            return body;
+        }
+
+        std::stringstream output;
+        output << "#include <stdexcept>\n";
+        output << "#include \"ps2_runtime_macros.h\"\n";
+        output << "#include \"ps2_runtime.h\"\n";
+        output << "#include \"ps2_syscalls.h\"\n";
+        output << "#include \"ps2_stubs.h\"\n\n";
+        output << "#ifdef PS2_FUNCTION_LOG_TRACKER\n";
+        output << "#include \"ps2_log.h\"\n";
+        output << "#endif\n\n";
+
+        // Most guest calls are dispatched through PS2Runtime and therefore do
+        // not need declarations for every generated function. A direct J to
+        // a recovered function is emitted as a C++ tail call, so declare only
+        // those targets locally instead of including the global declaration
+        // header in every generated source file.
+        constexpr std::string_view callSuffix = "(rdram, ctx, runtime); return;";
+        std::unordered_set<std::string> directJumpTargets;
+        size_t searchOffset = 0;
+        while ((searchOffset = body.find(callSuffix, searchOffset)) != std::string::npos)
+        {
+            const size_t lineStart = body.rfind('\n', searchOffset);
+            const size_t nameStart = body.find_first_not_of(" \t", lineStart == std::string::npos ? 0 : lineStart + 1);
+            if (nameStart != std::string::npos && nameStart < searchOffset)
+            {
+                directJumpTargets.emplace(body.substr(nameStart, searchOffset - nameStart));
+            }
+            searchOffset += callSuffix.size();
+        }
+        for (const std::string &target : directJumpTargets)
+        {
+            output << "void " << target << "(uint8_t*, R5900Context*, PS2Runtime*);\n";
+        }
+        if (!directJumpTargets.empty())
+        {
+            output << "\n";
+        }
+        output << body;
+        return output.str();
     }
 }
diff --git a/ps2xRecomp/src/lib/ps2_recompiler.cpp b/ps2xRecomp/src/lib/ps2_recompiler.cpp
index ab779a4f5..0f8a0fc03 100644
--- a/ps2xRecomp/src/lib/ps2_recompiler.cpp
+++ b/ps2xRecomp/src/lib/ps2_recompiler.cpp
@@ -102,7 +102,6 @@ namespace ps2recomp
         void writeCombinedOutputPreamble(std::ostream &output)
         {
             output << "#include <stdexcept>\n";
-            output << "#include \"ps2_recompiled_functions.h\"\n\n";
             output << "#include \"ps2_runtime_macros.h\"\n";
             output << "#include \"ps2_runtime.h\"\n";
             output << "#include \"ps2_recompiled_stubs.h\"\n";
diff --git a/ps2xTest/src/code_generator_tests.cpp b/ps2xTest/src/code_generator_tests.cpp
index 1a1f37025..da1c7319c 100644
--- a/ps2xTest/src/code_generator_tests.cpp
+++ b/ps2xTest/src/code_generator_tests.cpp
@@ -168,6 +168,48 @@ void register_code_generator_tests()
 {
     MiniTest::Case("CodeGenerator", [](TestCase &tc)
                    {
+    tc.Run("function sources avoid the global generated declaration header", [](TestCase &t) {
+        Function func;
+        func.name = "isolated_function";
+        func.start = 0x8000;
+        func.end = 0x8004;
+        func.isRecompiled = true;
+
+        CodeGenerator gen({}, {});
+        const std::string generated = gen.generateFunction(func, {makeNop(0x8000)}, true);
+
+        t.IsTrue(generated.find("#include \"ps2_runtime.h\"") != std::string::npos,
+                 "generated functions still require the runtime interface");
+        t.IsTrue(generated.find("#include \"ps2_recompiled_functions.h\"") == std::string::npos,
+                 "only the registration unit should depend on all generated declarations");
+        t.IsTrue(generated.find("#include \"ps2_recompiled_stubs.h\"") == std::string::npos,
+                 "ordinary generated functions should not depend on every stub declaration");
+    });
+
+    tc.Run("direct generated jumps declare only their target", [](TestCase &t) {
+        Function func;
+        func.name = "direct_jump_source";
+        func.start = 0x8000;
+        func.end = 0x8008;
+        func.isRecompiled = true;
+
+        Instruction jump{};
+        jump.address = 0x8000;
+        jump.opcode = OPCODE_J;
+        jump.target = (0x9000u >> 2) & 0x3FFFFFFu;
+        jump.hasDelaySlot = true;
+        jump.raw = (OPCODE_J << 26) | jump.target;
+
+        CodeGenerator gen({}, {});
+        gen.setRenamedFunctions({{0x9000u, "direct_jump_target"}});
+        const std::string generated = gen.generateFunction(func, {jump, makeNop(0x8004)}, true);
+
+        t.IsTrue(generated.find("void direct_jump_target(uint8_t*, R5900Context*, PS2Runtime*);") != std::string::npos,
+                 "a direct generated jump should receive a local forward declaration");
+        t.IsTrue(generated.find("#include \"ps2_recompiled_functions.h\"") == std::string::npos,
+                 "a direct generated jump should not restore the global declaration dependency");
+    });
+
     tc.Run("SYSCALL publishes its continuation before entering the runtime", [](TestCase &t) {
         Function func;
         func.name = "syscall_resume";