diff --git a/arm-gen.c b/arm-gen.c
index f535a09..634cffb 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -1410,6 +1410,16 @@ void gfunc_epilog(void)
   }
 }
 
+ST_FUNC void gen_fill_nops(int bytes)
+{
+    if ((bytes & 3))
+      tcc_error("alignment of code section not multiple of 4");
+    while (bytes > 0) {
+	o(0xE1A00000);
+	bytes -= 4;
+    }
+}
+
 /* generate a jump to a label */
 int gjmp(int t)
 {
diff --git a/arm64-gen.c b/arm64-gen.c
index 86b3af7..26cf5d1 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -1276,6 +1276,16 @@ ST_FUNC void gfunc_epilog(void)
     o(0xd65f03c0); // ret
 }
 
+ST_FUNC void gen_fill_nops(int bytes)
+{
+    if ((bytes & 3))
+      tcc_error("alignment of code section not multiple of 4");
+    while (bytes > 0) {
+	o(0xd503201f); // nop
+	bytes -= 4;
+    }
+}
+
 // Generate forward branch to label:
 ST_FUNC int gjmp(int t)
 {
diff --git a/c67-gen.c b/c67-gen.c
index bcb4b0e..d21626a 100644
--- a/c67-gen.c
+++ b/c67-gen.c
@@ -2035,6 +2035,16 @@ void gfunc_epilog(void)
     }
 }
 
+ST_FUNC void gen_fill_nops(int bytes)
+{
+    if ((bytes & 3))
+      tcc_error("alignment of code section not multiple of 4");
+    while (bytes > 0) {
+	C67_NOP(4);
+	bytes -= 4;
+    }
+}
+
 /* generate a jump to a label */
 int gjmp(int t)
 {
diff --git a/i386-gen.c b/i386-gen.c
index b9c3599..b6629d4 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -157,6 +157,12 @@ static int oad(int c, int s)
     return t;
 }
 
+ST_FUNC void gen_fill_nops(int bytes)
+{
+    while (bytes--)
+      g(0x90);
+}
+
 /* generate jmp to a label */
 #define gjmp2(instr,lbl) oad(instr,lbl)
 
diff --git a/tcc.h b/tcc.h
index cd67973..c2aaa42 100644
--- a/tcc.h
+++ b/tcc.h
@@ -1475,6 +1475,7 @@ ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *align, int *reg
 ST_FUNC void gfunc_call(int nb_args);
 ST_FUNC void gfunc_prolog(CType *func_type);
 ST_FUNC void gfunc_epilog(void);
+ST_FUNC void gen_fill_nops(int);
 ST_FUNC int gjmp(int t);
 ST_FUNC void gjmp_addr(int a);
 ST_FUNC int gtst(int inv, int t);
diff --git a/tccgen.c b/tccgen.c
index 6ef40e4..4f2e68b 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -7088,6 +7088,13 @@ static void gen_function(Sym *sym)
 {
     nocode_wanted = 0;
     ind = cur_text_section->data_offset;
+    if (sym->a.aligned) {
+	size_t newoff = section_add(cur_text_section, 0,
+				    1 << (sym->a.aligned - 1));
+	if (ind != newoff)
+	  gen_fill_nops(newoff - ind);
+	ind = newoff;
+    }
     /* NOTE: we patch the symbol size later */
     put_extern_sym(sym, cur_text_section, ind, 0);
     funcname = get_tok_str(sym->v, NULL);
diff --git a/tests/tcctest.c b/tests/tcctest.c
index 56d51dc..3dc9f1d 100644
--- a/tests/tcctest.c
+++ b/tests/tcctest.c
@@ -2270,6 +2270,8 @@ int fib(int n)
         return fib(n-1) + fib(n-2);
 }
 
+void __attribute__((aligned(16))) aligned_function(int i) {}
+
 void funcptr_test()
 {
     void (*func)(int);
@@ -2300,6 +2302,10 @@ void funcptr_test()
     func(42);
     (func + diff)(42);
     (num + a)(43);
+
+    /* Check that we can align functions */
+    func = aligned_function;
+    printf("aligned_function (should be zero): %d\n", ((int)func) & 15);
 }
 
 void lloptest(long long a, long long b)
diff --git a/x86_64-gen.c b/x86_64-gen.c
index e33a38a..9ca5c11 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -1639,6 +1639,12 @@ void gfunc_epilog(void)
 
 #endif /* not PE */
 
+ST_FUNC void gen_fill_nops(int bytes)
+{
+    while (bytes--)
+      g(0x90);
+}
+
 /* generate a jump to a label */
 int gjmp(int t)
 {