diff --git a/CMakeLists.txt b/CMakeLists.txt
index b4396fd4b3..60902c5e6d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -334,9 +334,15 @@ else()
 endif()
 
 if(WIN32)
-  list(APPEND MOLD_ELF_TEMPLATE_FILES src/subprocess-win32.cc)
+  list(APPEND MOLD_ELF_TEMPLATE_FILES
+    src/output-file-win32.cc
+    src/subprocess-win32.cc
+    )
 else()
-  list(APPEND MOLD_ELF_TEMPLATE_FILES src/subprocess-unix.cc)
+  list(APPEND MOLD_ELF_TEMPLATE_FILES
+    src/output-file-unix.cc
+    src/subprocess-unix.cc
+  )
 endif()
 
 function(mold_instantiate_templates SOURCE TARGET)
diff --git a/lib/common.h b/lib/common.h
index c27072410d..cfa9e9b547 100644
--- a/lib/common.h
+++ b/lib/common.h
@@ -525,15 +525,6 @@ inline void overwrite_uleb(u8 *loc, u64 val) {
   *loc = val & 0b0111'1111;
 }
 
-template <typename Context>
-std::string_view save_string(Context &ctx, const std::string &str) {
-  u8 *buf = new u8[str.size() + 1];
-  memcpy(buf, str.data(), str.size());
-  buf[str.size()] = '\0';
-  ctx.string_pool.push_back(std::unique_ptr<u8[]>(buf));
-  return {(char *)buf, str.size()};
-}
-
 static inline void pause() {
 #if defined(__x86_64__)
   asm volatile("pause");
@@ -718,82 +709,6 @@ class ConcurrentMap {
 
 void get_random_bytes(u8 *buf, i64 size);
 
-//
-// output-file.h
-//
-
-template <typename Context>
-class OutputFile {
-public:
-  static std::unique_ptr<OutputFile<Context>>
-  open(Context &ctx, std::string path, i64 filesize, int perm);
-
-  virtual void close(Context &ctx) = 0;
-  virtual ~OutputFile() = default;
-
-  u8 *buf = nullptr;
-  std::vector<u8> buf2;
-  std::string path;
-  int fd = -1;
-  i64 filesize = 0;
-  bool is_mmapped = false;
-  bool is_unmapped = false;
-
-protected:
-  OutputFile(std::string path, i64 filesize, bool is_mmapped)
-    : path(path), filesize(filesize), is_mmapped(is_mmapped) {}
-};
-
-template <typename Context>
-class MallocOutputFile : public OutputFile<Context> {
-public:
-  MallocOutputFile(Context &ctx, std::string path, i64 filesize, int perm)
-    : OutputFile<Context>(path, filesize, false), ptr(new u8[filesize]),
-      perm(perm) {
-    this->buf = ptr.get();
-  }
-
-  void close(Context &ctx) override {
-    Timer t(ctx, "close_file");
-    FILE *fp;
-
-    if (this->path == "-") {
-      fp = stdout;
-    } else {
-#ifdef _WIN32
-      int pmode = (perm & 0200) ? (_S_IREAD | _S_IWRITE) : _S_IREAD;
-      i64 fd = _open(this->path.c_str(), _O_RDWR | _O_CREAT | _O_BINARY, pmode);
-#else
-      i64 fd = ::open(this->path.c_str(), O_RDWR | O_CREAT, perm);
-#endif
-      if (fd == -1)
-        Fatal(ctx) << "cannot open " << this->path << ": " << errno_string();
-#ifdef _WIN32
-      fp = _fdopen(fd, "wb");
-#else
-      fp = fdopen(fd, "w");
-#endif
-    }
-
-    fwrite(this->buf, this->filesize, 1, fp);
-    if (!this->buf2.empty())
-      fwrite(this->buf2.data(), this->buf2.size(), 1, fp);
-    fclose(fp);
-  }
-
-private:
-  std::unique_ptr<u8[]> ptr;
-  int perm;
-};
-
-template <typename Context>
-class LockingOutputFile : public OutputFile<Context> {
-public:
-  LockingOutputFile(Context &ctx, std::string path, int perm);
-  void resize(Context &ctx, i64 filesize);
-  void close(Context &ctx) override;
-};
-
 //
 // hyperloglog.cc
 //
diff --git a/lib/output-file.h b/lib/output-file.h
deleted file mode 100644
index 63299ed998..0000000000
--- a/lib/output-file.h
+++ /dev/null
@@ -1,5 +0,0 @@
-#if _WIN32
-# include "output-file-win32.h"
-#else
-# include "output-file-unix.h"
-#endif
diff --git a/src/main.cc b/src/main.cc
index 2dc4c8235b..0115c596c3 100644
--- a/src/main.cc
+++ b/src/main.cc
@@ -1,7 +1,6 @@
 #include "mold.h"
 #include "filetype.h"
 #include "../lib/archive-file.h"
-#include "../lib/output-file.h"
 
 #include <cstring>
 #include <functional>
@@ -655,8 +654,7 @@ int mold_main(int argc, char **argv) {
   t_before_copy.stop();
 
   // Create an output file
-  ctx.output_file =
-    OutputFile<Context<E>>::open(ctx, ctx.arg.output, filesize, 0777);
+  ctx.output_file = OutputFile<E>::open(ctx, ctx.arg.output, filesize, 0777);
   ctx.buf = ctx.output_file->buf;
 
   Timer t_copy(ctx, "copy");
diff --git a/src/mold.h b/src/mold.h
index 4149f045ba..b95710dfdf 100644
--- a/src/mold.h
+++ b/src/mold.h
@@ -1074,6 +1074,82 @@ class ComdatGroupSection : public Chunk<E> {
   std::vector<Chunk<E> *> members;
 };
 
+//
+// output-file.cc
+//
+
+template <typename E>
+class OutputFile {
+public:
+  static std::unique_ptr<OutputFile<E>>
+  open(Context<E> &ctx, std::string path, i64 filesize, int perm);
+
+  virtual void close(Context<E> &ctx) = 0;
+  virtual ~OutputFile() = default;
+
+  u8 *buf = nullptr;
+  std::vector<u8> buf2;
+  std::string path;
+  int fd = -1;
+  i64 filesize = 0;
+  bool is_mmapped = false;
+  bool is_unmapped = false;
+
+protected:
+  OutputFile(std::string path, i64 filesize, bool is_mmapped)
+    : path(path), filesize(filesize), is_mmapped(is_mmapped) {}
+};
+
+template <typename E>
+class MallocOutputFile : public OutputFile<E> {
+public:
+  MallocOutputFile(Context<E> &ctx, std::string path, i64 filesize, int perm)
+    : OutputFile<E>(path, filesize, false), ptr(new u8[filesize]),
+      perm(perm) {
+    this->buf = ptr.get();
+  }
+
+  void close(Context<E> &ctx) override {
+    Timer t(ctx, "close_file");
+    FILE *fp;
+
+    if (this->path == "-") {
+      fp = stdout;
+    } else {
+#ifdef _WIN32
+      int pmode = (perm & 0200) ? (_S_IREAD | _S_IWRITE) : _S_IREAD;
+      i64 fd = _open(this->path.c_str(), _O_RDWR | _O_CREAT | _O_BINARY, pmode);
+#else
+      i64 fd = ::open(this->path.c_str(), O_RDWR | O_CREAT, perm);
+#endif
+      if (fd == -1)
+        Fatal(ctx) << "cannot open " << this->path << ": " << errno_string();
+#ifdef _WIN32
+      fp = _fdopen(fd, "wb");
+#else
+      fp = fdopen(fd, "w");
+#endif
+    }
+
+    fwrite(this->buf, this->filesize, 1, fp);
+    if (!this->buf2.empty())
+      fwrite(this->buf2.data(), this->buf2.size(), 1, fp);
+    fclose(fp);
+  }
+
+private:
+  std::unique_ptr<u8[]> ptr;
+  int perm;
+};
+
+template <typename E>
+class LockingOutputFile : public OutputFile<E> {
+public:
+  LockingOutputFile(Context<E> &ctx, std::string path, int perm);
+  void resize(Context<E> &ctx, i64 filesize);
+  void close(Context<E> &ctx) override;
+};
+
 //
 // gdb-index.cc
 //
@@ -1956,7 +2032,7 @@ struct Context {
   std::vector<ElfSym<E>> internal_esyms;
 
   // Output buffer
-  std::unique_ptr<OutputFile<Context<E>>> output_file;
+  std::unique_ptr<OutputFile<E>> output_file;
   u8 *buf = nullptr;
   bool overwrite_output_file = true;
 
@@ -3021,4 +3097,13 @@ inline bool is_c_identifier(std::string_view s) {
   return true;
 }
 
+template <typename E>
+std::string_view save_string(Context<E> &ctx, const std::string &str) {
+  u8 *buf = new u8[str.size() + 1];
+  memcpy(buf, str.data(), str.size());
+  buf[str.size()] = '\0';
+  ctx.string_pool.push_back(std::unique_ptr<u8[]>(buf));
+  return {(char *)buf, str.size()};
+}
+
 } // namespace mold
diff --git a/lib/output-file-unix.h b/src/output-file-unix.cc
similarity index 83%
rename from lib/output-file-unix.h
rename to src/output-file-unix.cc
index dcbebfb3f0..0a6f9eb28a 100644
--- a/lib/output-file-unix.h
+++ b/src/output-file-unix.cc
@@ -1,4 +1,4 @@
-#include "common.h"
+#include "mold.h"
 
 #include <fcntl.h>
 #include <filesystem>
@@ -9,15 +9,15 @@
 
 namespace mold {
 
-inline u32 get_umask() {
+static u32 get_umask() {
   u32 orig_umask = umask(0);
   umask(orig_umask);
   return orig_umask;
 }
 
-template <typename Context>
+template <typename E>
 static int
-open_or_create_file(Context &ctx, std::string path, std::string tmpfile,
+open_or_create_file(Context<E> &ctx, std::string path, std::string tmpfile,
                     int perm) {
   // Reuse an existing file if exists and writable because on Linux,
   // writing to an existing file is much faster than creating a fresh
@@ -35,11 +35,11 @@ open_or_create_file(Context &ctx, std::string path, std::string tmpfile,
   return fd;
 }
 
-template <typename Context>
-class MemoryMappedOutputFile : public OutputFile<Context> {
+template <typename E>
+class MemoryMappedOutputFile : public OutputFile<E> {
 public:
-  MemoryMappedOutputFile(Context &ctx, std::string path, i64 filesize, int perm)
-    : OutputFile<Context>(path, filesize, true) {
+  MemoryMappedOutputFile(Context<E> &ctx, std::string path, i64 filesize, int perm)
+    : OutputFile<E>(path, filesize, true) {
     std::filesystem::path dir = filepath(path).parent_path();
     std::string filename = filepath(path).filename().string();
     std::string tmpfile = dir / ("." + filename + "." + std::to_string(getpid()));
@@ -72,7 +72,7 @@ class MemoryMappedOutputFile : public OutputFile<Context> {
       ::close(fd2);
   }
 
-  void close(Context &ctx) override {
+  void close(Context<E> &ctx) override {
     Timer t(ctx, "close_file");
 
     if (!this->is_unmapped)
@@ -103,9 +103,9 @@ class MemoryMappedOutputFile : public OutputFile<Context> {
   int fd2 = -1;
 };
 
-template <typename Context>
-std::unique_ptr<OutputFile<Context>>
-OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm) {
+template <typename E>
+std::unique_ptr<OutputFile<E>>
+OutputFile<E>::open(Context<E> &ctx, std::string path, i64 filesize, int perm) {
   Timer t(ctx, "open_file");
 
   if (path.starts_with('/') && !ctx.arg.chroot.empty())
@@ -120,7 +120,7 @@ OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm
       is_special = true;
   }
 
-  OutputFile<Context> *file;
+  OutputFile<E> *file;
   if (is_special)
     file = new MallocOutputFile(ctx, path, filesize, perm);
   else
@@ -146,10 +146,10 @@ OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm
 
 // LockingOutputFile is similar to MemoryMappedOutputFile, but it doesn't
 // rename output files and instead acquires file lock using flock().
-template <typename Context>
-LockingOutputFile<Context>::LockingOutputFile(Context &ctx, std::string path,
-                                              int perm)
-  : OutputFile<Context>(path, 0, true) {
+template <typename E>
+LockingOutputFile<E>::LockingOutputFile(Context<E> &ctx, std::string path,
+                                        int perm)
+  : OutputFile<E>(path, 0, true) {
   this->fd = ::open(path.c_str(), O_RDWR | O_CREAT, perm);
   if (this->fd == -1)
     Fatal(ctx) << "cannot open " << path << ": " << errno_string();
@@ -162,8 +162,8 @@ LockingOutputFile<Context>::LockingOutputFile(Context &ctx, std::string path,
   (void)!!write(this->fd, buf, sizeof(buf));
 }
 
-template <typename Context>
-void LockingOutputFile<Context>::resize(Context &ctx, i64 filesize) {
+template <typename E>
+void LockingOutputFile<E>::resize(Context<E> &ctx, i64 filesize) {
   if (ftruncate(this->fd, filesize) == -1)
     Fatal(ctx) << "ftruncate failed: " << errno_string();
 
@@ -177,8 +177,8 @@ void LockingOutputFile<Context>::resize(Context &ctx, i64 filesize) {
   mold::output_buffer_end = this->buf + filesize;
 }
 
-template <typename Context>
-void LockingOutputFile<Context>::close(Context &ctx) {
+template <typename E>
+void LockingOutputFile<E>::close(Context<E> &ctx) {
   if (!this->is_unmapped)
     munmap(this->buf, this->filesize);
 
@@ -192,4 +192,9 @@ void LockingOutputFile<Context>::close(Context &ctx) {
   ::close(this->fd);
 }
 
+using E = MOLD_TARGET;
+
+template class OutputFile<E>;
+template class LockingOutputFile<E>;
+
 } // namespace mold
diff --git a/lib/output-file-win32.h b/src/output-file-win32.cc
similarity index 74%
rename from lib/output-file-win32.h
rename to src/output-file-win32.cc
index 5e37fb47a1..68bd26c8a5 100644
--- a/lib/output-file-win32.h
+++ b/src/output-file-win32.cc
@@ -1,4 +1,4 @@
-#include "common.h"
+#include "mold.h"
 
 #include <fcntl.h>
 #include <filesystem>
@@ -6,11 +6,11 @@
 
 namespace mold {
 
-template <typename Context>
-class MemoryMappedOutputFile : public OutputFile<Context> {
+template <typename E>
+class MemoryMappedOutputFile : public OutputFile<E> {
 public:
-  MemoryMappedOutputFile(Context &ctx, std::string path, i64 filesize, int perm)
-      : OutputFile<Context>(path, filesize, true) {
+  MemoryMappedOutputFile(Context<E> &ctx, std::string path, i64 filesize, int perm)
+      : OutputFile<E>(path, filesize, true) {
     // TODO: use intermediate temporary file for output.
     DWORD attrs = (perm & 0200) ? FILE_ATTRIBUTE_NORMAL : FILE_ATTRIBUTE_READONLY;
 
@@ -40,7 +40,7 @@ class MemoryMappedOutputFile : public OutputFile<Context> {
       CloseHandle(handle);
   }
 
-  void close(Context &ctx) override {
+  void close(Context<E> &ctx) override {
     Timer t(ctx, "close_file");
 
     UnmapViewOfFile(this->buf);
@@ -64,9 +64,9 @@ class MemoryMappedOutputFile : public OutputFile<Context> {
   HANDLE handle;
 };
 
-template <typename Context>
-std::unique_ptr<OutputFile<Context>>
-OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm) {
+template <typename E>
+std::unique_ptr<OutputFile<E>>
+OutputFile<E>::open(Context<E> &ctx, std::string path, i64 filesize, int perm) {
   Timer t(ctx, "open_file");
 
   if (path.starts_with('/') && !ctx.arg.chroot.empty())
@@ -86,7 +86,7 @@ OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm
     }
   }
 
-  OutputFile<Context> *file;
+  OutputFile<E> *file;
   if (is_special)
     file = new MallocOutputFile(ctx, path, filesize, perm);
   else
@@ -94,20 +94,25 @@ OutputFile<Context>::open(Context &ctx, std::string path, i64 filesize, int perm
 
   if (ctx.arg.filler != -1)
     memset(file->buf, ctx.arg.filler, filesize);
-  return std::unique_ptr<OutputFile<Context>>(file);
+  return std::unique_ptr<OutputFile<E>>(file);
 }
 
-template <typename Context>
-LockingOutputFile<Context>::LockingOutputFile(Context &ctx, std::string path,
-                                              int perm)
-  : OutputFile<Context>(path, 0, true) {
+template <typename E>
+LockingOutputFile<E>::LockingOutputFile(Context<E> &ctx, std::string path,
+                                        int perm)
+  : OutputFile<E>(path, 0, true) {
   Fatal(ctx) << "LockingOutputFile is not supported on Windows";
 }
 
-template <typename Context>
-void LockingOutputFile<Context>::resize(Context &ctx, i64 filesize) {}
+template <typename E>
+void LockingOutputFile<E>::resize(Context<E> &ctx, i64 filesize) {}
 
-template <typename Context>
-void LockingOutputFile<Context>::close(Context &ctx) {}
+template <typename E>
+void LockingOutputFile<E>::close(Context<E> &ctx) {}
+
+using E = MOLD_TARGET;
+
+template class OutputFile<E>;
+template class LockingOutputFile<E>;
 
 } // namespace mold
diff --git a/src/passes.cc b/src/passes.cc
index 24b2c35984..099dffcaed 100644
--- a/src/passes.cc
+++ b/src/passes.cc
@@ -1,6 +1,5 @@
 #include "mold.h"
 #include "blake3.h"
-#include "../lib/output-file.h"
 
 #include <fstream>
 #include <functional>
@@ -3076,8 +3075,8 @@ void write_separate_debug_file(Context<E> &ctx) {
   Timer t(ctx, "write_separate_debug_file");
 
   // Open an output file early
-  LockingOutputFile<Context<E>> *file =
-    new LockingOutputFile<Context<E>>(ctx, ctx.arg.separate_debug_file, 0666);
+  LockingOutputFile<E> *file =
+    new LockingOutputFile<E>(ctx, ctx.arg.separate_debug_file, 0666);
 
   // We want to write to the debug info file in background so that the
   // user doesn't have to wait for it to complete.
diff --git a/src/relocatable.cc b/src/relocatable.cc
index d582c2edc7..639dc6ae87 100644
--- a/src/relocatable.cc
+++ b/src/relocatable.cc
@@ -169,8 +169,7 @@ void combine_objects(Context<E> &ctx) {
   compute_section_headers(ctx);
 
   i64 filesize = r_set_osec_offsets(ctx);
-  ctx.output_file =
-    OutputFile<Context<E>>::open(ctx, ctx.arg.output, filesize, 0666);
+  ctx.output_file = OutputFile<E>::open(ctx, ctx.arg.output, filesize, 0666);
   ctx.buf = ctx.output_file->buf;
 
   copy_chunks(ctx);