diff --git a/include/tccdefs.h b/include/tccdefs.h
index eaf0c58c..5c6535a0 100644
--- a/include/tccdefs.h
+++ b/include/tccdefs.h
@@ -191,7 +191,10 @@
 #if defined __x86_64__
 #if !defined _WIN32
     /* GCC compatible definition of va_list. */
-    /* This should be in sync with the declaration in our lib/libtcc1.c */
+
+    enum __va_arg_type {
+        __va_gen_reg, __va_float_reg, __va_stack
+    };
     typedef struct {
         unsigned gp_offset, fp_offset;
         union {
@@ -201,7 +204,43 @@
         char *reg_save_area;
     } __builtin_va_list[1];
 
-    void *__va_arg(__builtin_va_list ap, int arg_type, int size, int align);
+    static inline void *__va_arg(__builtin_va_list ap, int arg_type,
+                                 int size, int align)
+    {
+        size = (size + 7) & ~7;
+        align = (align + 7) & ~7;
+        switch ((enum __va_arg_type)arg_type) {
+        case __va_gen_reg:
+            if (ap->gp_offset + size <= 48) {
+                ap->gp_offset += size;
+                return ap->reg_save_area + ap->gp_offset - size;
+            }
+            goto use_overflow_area;
+        case __va_float_reg:
+            if (ap->fp_offset < 128 + 48) {
+                ap->fp_offset += 16;
+                if (size == 8)
+                    return ap->reg_save_area + ap->fp_offset - 16;
+                if (ap->fp_offset < 128 + 48) {
+                    double *p = (double *)(ap->reg_save_area + ap->fp_offset);
+                    p[-1] = p[0];
+                    ap->fp_offset += 16;
+                    return ap->reg_save_area + ap->fp_offset - 32;
+                }
+            }
+            goto use_overflow_area;
+        case __va_stack:
+        use_overflow_area:
+            ap->overflow_arg_area += size;
+            ap->overflow_arg_area =
+	      (char*)((long long)(ap->overflow_arg_area + align - 1) & -align);
+            return ap->overflow_arg_area - size;
+        default: /* should never happen */
+            char *a = (char *)0; *a = 0; // abort
+            return 0;
+        }
+    }
+
     #define __builtin_va_start(ap, last) \
        (*(ap) = *(__builtin_va_list)((char*)__builtin_frame_address(0) - 24))
     #define __builtin_va_arg(ap, t)   \
diff --git a/lib/Makefile b/lib/Makefile
index 29b339a2..26d23d3a 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -59,8 +59,8 @@ $(Cbc)COMMON_O += bcheck.o
 EXTRA_O = runmain.o bt-exe.o bt-dll.o bt-log.o bcheck.o
 
 OBJ-i386 = $(I386_O) $(LIN_O)
-OBJ-x86_64 = $(X86_64_O) va_list.o $(LIN_O)
-OBJ-x86_64-osx = $(X86_64_O) va_list.o $(OSX_O)
+OBJ-x86_64 = $(X86_64_O) $(LIN_O)
+OBJ-x86_64-osx = $(X86_64_O) $(OSX_O)
 OBJ-i386-win32 = $(I386_O) chkstk.o $(WIN_O)
 OBJ-x86_64-win32 = $(X86_64_O) chkstk.o $(WIN_O)
 OBJ-arm64 = $(ARM64_O) $(LIN_O)
diff --git a/lib/va_list.c b/lib/va_list.c
deleted file mode 100644
index 1fb55127..00000000
--- a/lib/va_list.c
+++ /dev/null
@@ -1,67 +0,0 @@
-/* va_list.c - tinycc support for va_list on X86_64 */
-
-#if defined __x86_64__
-
-/* Avoid include files, they may not be available when cross compiling */
-extern void abort(void);
-
-/* This should be in sync with our include/stdarg.h */
-enum __va_arg_type {
-    __va_gen_reg, __va_float_reg, __va_stack
-};
-
-/* GCC compatible definition of va_list. */
-/*predefined by TCC (tcc_predefs.h):
-typedef struct {
-    unsigned int gp_offset;
-    unsigned int fp_offset;
-    union {
-        unsigned int overflow_offset;
-        char *overflow_arg_area;
-    };
-    char *reg_save_area;
-} __builtin_va_list[1];
-*/
-
-extern void *memcpy(void *dest, const void *src, unsigned long n);
-
-void *__va_arg(__builtin_va_list ap,
-               int arg_type,
-               int size, int align)
-{
-    size = (size + 7) & ~7;
-    align = (align + 7) & ~7;
-    switch ((enum __va_arg_type)arg_type) {
-    case __va_gen_reg:
-        if (ap->gp_offset + size <= 48) {
-            ap->gp_offset += size;
-            return ap->reg_save_area + ap->gp_offset - size;
-        }
-        goto use_overflow_area;
-
-    case __va_float_reg:
-        if (ap->fp_offset < 128 + 48) {
-            ap->fp_offset += 16;
-            if (size == 8)
-                return ap->reg_save_area + ap->fp_offset - 16;
-            if (ap->fp_offset < 128 + 48) {
-                memcpy(ap->reg_save_area + ap->fp_offset - 8,
-                       ap->reg_save_area + ap->fp_offset, 8);
-                ap->fp_offset += 16;
-                return ap->reg_save_area + ap->fp_offset - 32;
-            }
-        }
-        goto use_overflow_area;
-
-    case __va_stack:
-    use_overflow_area:
-        ap->overflow_arg_area += size;
-        ap->overflow_arg_area = (char*)((long long)(ap->overflow_arg_area + align - 1) & -align);
-        return ap->overflow_arg_area - size;
-
-    default: /* should never happen */
-        abort();
-        return 0;
-    }
-}
-#endif