py: Implement keyword-only args.

Implements 'def f(*, a)' and 'def f(*a, b)', but not default
keyword-only args, eg 'def f(*, a=1)'.

Partially addresses issue #524.

12 files changed, 198 insertions, 70 deletions

diff --git a/py/compile.c b/py/compile.c
index 532e29a233..2580cbeb51 100644
--- a/py/compile.c
+++ b/py/compile.c
@@ -2789,14 +2789,17 @@ void compile_node(compiler_t *comp, mp_parse_node_t pn) {
 
 void compile_scope_func_lambda_param(compiler_t *comp, mp_parse_node_t pn, pn_kind_t pn_name, pn_kind_t pn_star, pn_kind_t pn_dbl_star, bool allow_annotations) {
     // TODO verify that *k and **k are last etc
-    qstr param_name = 0;
+    qstr param_name = MP_QSTR_NULL;
+    uint param_flag = ID_FLAG_IS_PARAM;
     mp_parse_node_t pn_annotation = MP_PARSE_NODE_NULL;
     if (MP_PARSE_NODE_IS_ID(pn)) {
         param_name = MP_PARSE_NODE_LEAF_ARG(pn);
         if (comp->have_star) {
-            // comes after a bare star, so doesn't count as a parameter
+            // comes after a star, so counts as a keyword-only parameter
+            comp->scope_cur->num_kwonly_args += 1;
         } else {
-            comp->scope_cur->num_params += 1;
+            // comes before a star, so counts as a positional parameter
+            comp->scope_cur->num_pos_args += 1;
         }
     } else {
         assert(MP_PARSE_NODE_IS_STRUCT(pn));
@@ -2822,12 +2825,15 @@ void compile_scope_func_lambda_param(compiler_t *comp, mp_parse_node_t pn, pn_ki
             }
             */
             if (comp->have_star) {
-                // comes after a bare star, so doesn't count as a parameter
+                // comes after a star, so counts as a keyword-only parameter
+                comp->scope_cur->num_kwonly_args += 1;
             } else {
-                comp->scope_cur->num_params += 1;
+                // comes before a star, so counts as a positional parameter
+                comp->scope_cur->num_pos_args += 1;
             }
         } else if (MP_PARSE_NODE_STRUCT_KIND(pns) == pn_star) {
             comp->have_star = true;
+            param_flag = ID_FLAG_IS_PARAM | ID_FLAG_IS_STAR_PARAM;
             if (MP_PARSE_NODE_IS_NULL(pns->nodes[0])) {
                 // bare star
                 // TODO see http://www.python.org/dev/peps/pep-3102/
@@ -2848,6 +2854,7 @@ void compile_scope_func_lambda_param(compiler_t *comp, mp_parse_node_t pn, pn_ki
             }
         } else if (MP_PARSE_NODE_STRUCT_KIND(pns) == pn_dbl_star) {
             param_name = MP_PARSE_NODE_LEAF_ARG(pns->nodes[0]);
+            param_flag = ID_FLAG_IS_PARAM | ID_FLAG_IS_DBL_STAR_PARAM;
             if (allow_annotations && !MP_PARSE_NODE_IS_NULL(pns->nodes[1])) {
                 // this parameter has an annotation
                 pn_annotation = pns->nodes[1];
@@ -2859,7 +2866,7 @@ void compile_scope_func_lambda_param(compiler_t *comp, mp_parse_node_t pn, pn_ki
         }
     }
 
-    if (param_name != 0) {
+    if (param_name != MP_QSTR_NULL) {
         if (!MP_PARSE_NODE_IS_NULL(pn_annotation)) {
             // TODO this parameter has an annotation
         }
@@ -2870,15 +2877,15 @@ void compile_scope_func_lambda_param(compiler_t *comp, mp_parse_node_t pn, pn_ki
             return;
         }
         id_info->kind = ID_INFO_KIND_LOCAL;
-        id_info->flags |= ID_FLAG_IS_PARAM;
+        id_info->flags = param_flag;
     }
 }
 
-void compile_scope_func_param(compiler_t *comp, mp_parse_node_t pn) {
+STATIC void compile_scope_func_param(compiler_t *comp, mp_parse_node_t pn) {
     compile_scope_func_lambda_param(comp, pn, PN_typedargslist_name, PN_typedargslist_star, PN_typedargslist_dbl_star, true);
 }
 
-void compile_scope_lambda_param(compiler_t *comp, mp_parse_node_t pn) {
+STATIC void compile_scope_lambda_param(compiler_t *comp, mp_parse_node_t pn) {
     compile_scope_func_lambda_param(comp, pn, PN_varargslist_name, PN_varargslist_star, PN_varargslist_dbl_star, false);
 }
 
@@ -3051,7 +3058,7 @@ STATIC void compile_scope(compiler_t *comp, scope_t *scope, pass_kind_t pass) {
             id_info_t *id_info = scope_find_or_add_id(comp->scope_cur, qstr_arg, &added);
             assert(added);
             id_info->kind = ID_INFO_KIND_LOCAL;
-            scope->num_params = 1;
+            scope->num_pos_args = 1;
         }
 
         if (scope->kind == SCOPE_LIST_COMP) {
@@ -3144,7 +3151,7 @@ STATIC void compile_scope_inline_asm(compiler_t *comp, scope_t *scope, pass_kind
     if (comp->pass == PASS_2) {
         mp_parse_node_t *pn_params;
         int n_params = list_get(&pns->nodes[1], PN_typedargslist, &pn_params);
-        scope->num_params = EMIT_INLINE_ASM_ARG(count_params, n_params, pn_params);
+        scope->num_pos_args = EMIT_INLINE_ASM_ARG(count_params, n_params, pn_params);
     }
 
     assert(MP_PARSE_NODE_IS_NULL(pns->nodes[2])); // type
@@ -3235,6 +3242,25 @@ STATIC void compile_scope_inline_asm(compiler_t *comp, scope_t *scope, pass_kind
 #endif
 
 STATIC void compile_scope_compute_things(compiler_t *comp, scope_t *scope) {
+#if !MICROPY_EMIT_CPYTHON
+    // in Micro Python we put the *x parameter after all other parameters (except **y)
+    if (scope->scope_flags & MP_SCOPE_FLAG_VARARGS) {
+        id_info_t *id_param = NULL;
+        for (int i = scope->id_info_len - 1; i >= 0; i--) {
+            id_info_t *id = &scope->id_info[i];
+            if (id->flags & ID_FLAG_IS_STAR_PARAM) {
+                if (id_param != NULL) {
+                    // swap star param with last param
+                    id_info_t temp = *id_param; *id_param = *id; *id = temp;
+                }
+                break;
+            } else if (id_param == NULL && id->flags == ID_FLAG_IS_PARAM) {
+                id_param = id;
+            }
+        }
+    }
+#endif
+
     // in functions, turn implicit globals into explicit globals
     // compute the index of each local
     scope->num_locals = 0;
@@ -3247,10 +3273,9 @@ STATIC void compile_scope_compute_things(compiler_t *comp, scope_t *scope) {
         if (scope->kind >= SCOPE_FUNCTION && scope->kind <= SCOPE_GEN_EXPR && id->kind == ID_INFO_KIND_GLOBAL_IMPLICIT) {
             id->kind = ID_INFO_KIND_GLOBAL_EXPLICIT;
         }
-        // note: params always count for 1 local, even if they are a cell
+        // params always count for 1 local, even if they are a cell
         if (id->kind == ID_INFO_KIND_LOCAL || (id->flags & ID_FLAG_IS_PARAM)) {
-            id->local_num = scope->num_locals;
-            scope->num_locals += 1;
+            id->local_num = scope->num_locals++;
         }
     }
 
@@ -3309,7 +3334,7 @@ STATIC void compile_scope_compute_things(compiler_t *comp, scope_t *scope) {
                     id->local_num += num_free;
                 }
             }
-            scope->num_params += num_free; // free vars are counted as params for passing them into the function
+            scope->num_pos_args += num_free; // free vars are counted as params for passing them into the function
             scope->num_locals += num_free;
         }
 #endif
diff --git a/py/emitbc.c b/py/emitbc.c
index f58cec1f1f..3e3f0751ca 100644
--- a/py/emitbc.c
+++ b/py/emitbc.c
@@ -300,14 +300,14 @@ STATIC void emit_bc_end_pass(emit_t *emit) {
         emit->code_base = m_new0(byte, emit->code_info_size + emit->byte_code_size);
 
     } else if (emit->pass == PASS_3) {
-        qstr *arg_names = m_new(qstr, emit->scope->num_params);
-        for (int i = 0; i < emit->scope->num_params; i++) {
+        qstr *arg_names = m_new(qstr, emit->scope->num_pos_args + emit->scope->num_kwonly_args);
+        for (int i = 0; i < emit->scope->num_pos_args + emit->scope->num_kwonly_args; i++) {
             arg_names[i] = emit->scope->id_info[i].qstr;
         }
         mp_emit_glue_assign_byte_code(emit->scope->raw_code, emit->code_base,
             emit->code_info_size + emit->byte_code_size,
-            emit->scope->num_params, emit->scope->num_locals,
-            emit->scope->scope_flags, arg_names);
+            emit->scope->num_pos_args, emit->scope->num_kwonly_args, arg_names,
+            emit->scope->scope_flags);
     }
 }
 
diff --git a/py/emitglue.c b/py/emitglue.c
index dddeea1b57..d39f554a3b 100644
--- a/py/emitglue.c
+++ b/py/emitglue.c
@@ -47,16 +47,22 @@ mp_raw_code_t *mp_emit_glue_new_raw_code(void) {
     return rc;
 }
 
-void mp_emit_glue_assign_byte_code(mp_raw_code_t *rc, byte *code, uint len, int n_args, int n_locals, uint scope_flags, qstr *arg_names) {
+void mp_emit_glue_assign_byte_code(mp_raw_code_t *rc, byte *code, uint len, uint n_pos_args, uint n_kwonly_args, qstr *arg_names, uint scope_flags) {
     rc->kind = MP_CODE_BYTE;
     rc->scope_flags = scope_flags;
-    rc->n_args = n_args;
+    rc->n_pos_args = n_pos_args;
+    rc->n_kwonly_args = n_kwonly_args;
+    rc->arg_names = arg_names;
     rc->u_byte.code = code;
     rc->u_byte.len = len;
-    rc->arg_names = arg_names;
 
 #ifdef DEBUG_PRINT
-    DEBUG_printf("assign byte code: code=%p len=%u n_args=%d n_locals=%d\n", code, len, n_args, n_locals);
+    DEBUG_printf("assign byte code: code=%p len=%u n_pos_args=%d n_kwonly_args=%d flags=%x\n", code, len, n_pos_args, n_kwonly_args, scope_flags);
+    DEBUG_printf("  arg names:");
+    for (int i = 0; i < n_pos_args + n_kwonly_args; i++) {
+        DEBUG_printf(" %s", qstr_str(arg_names[i]));
+    }
+    DEBUG_printf("\n");
     for (int i = 0; i < 128 && i < len; i++) {
         if (i > 0 && i % 16 == 0) {
             DEBUG_printf("\n");
@@ -73,7 +79,7 @@ void mp_emit_glue_assign_byte_code(mp_raw_code_t *rc, byte *code, uint len, int
 void mp_emit_glue_assign_native_code(mp_raw_code_t *rc, void *fun, uint len, int n_args) {
     rc->kind = MP_CODE_NATIVE;
     rc->scope_flags = 0;
-    rc->n_args = n_args;
+    rc->n_pos_args = n_args;
     rc->u_native.fun = fun;
 
 #ifdef DEBUG_PRINT
@@ -99,7 +105,7 @@ void mp_emit_glue_assign_native_code(mp_raw_code_t *rc, void *fun, uint len, int
 void mp_emit_glue_assign_inline_asm_code(mp_raw_code_t *rc, void *fun, uint len, int n_args) {
     rc->kind = MP_CODE_INLINE_ASM;
     rc->scope_flags = 0;
-    rc->n_args = n_args;
+    rc->n_pos_args = n_args;
     rc->u_inline_asm.fun = fun;
 
 #ifdef DEBUG_PRINT
@@ -136,13 +142,13 @@ mp_obj_t mp_make_function_from_raw_code(mp_raw_code_t *rc, mp_obj_t def_args, mp
     mp_obj_t fun;
     switch (rc->kind) {
         case MP_CODE_BYTE:
-            fun = mp_obj_new_fun_bc(rc->scope_flags, rc->arg_names, rc->n_args, def_args, rc->u_byte.code);
+            fun = mp_obj_new_fun_bc(rc->scope_flags, rc->arg_names, rc->n_pos_args, rc->n_kwonly_args, def_args, rc->u_byte.code);
             break;
         case MP_CODE_NATIVE:
-            fun = mp_make_function_n(rc->n_args, rc->u_native.fun);
+            fun = mp_make_function_n(rc->n_pos_args, rc->u_native.fun);
             break;
         case MP_CODE_INLINE_ASM:
-            fun = mp_obj_new_fun_asm(rc->n_args, rc->u_inline_asm.fun);
+            fun = mp_obj_new_fun_asm(rc->n_pos_args, rc->u_inline_asm.fun);
             break;
         default:
             // raw code was never set (this should not happen)
diff --git a/py/emitglue.h b/py/emitglue.h
index 9aea2e4d45..f15e5b7161 100644
--- a/py/emitglue.h
+++ b/py/emitglue.h
@@ -9,9 +9,11 @@ typedef enum {
 } mp_raw_code_kind_t;
 
 typedef struct _mp_code_t {
-    mp_raw_code_kind_t kind : 8;
-    uint scope_flags : 8;
-    uint n_args : 16;
+    mp_raw_code_kind_t kind : 3;
+    uint scope_flags : 7;
+    uint n_pos_args : 11;
+    uint n_kwonly_args : 11;
+    qstr *arg_names;
     union {
         struct {
             byte *code;
@@ -24,7 +26,6 @@ typedef struct _mp_code_t {
             void *fun;
         } u_inline_asm;
     };
-    qstr *arg_names;
 } mp_raw_code_t;
 
 void mp_emit_glue_init(void);
@@ -32,7 +33,7 @@ void mp_emit_glue_deinit(void);
 
 mp_raw_code_t *mp_emit_glue_new_raw_code(void);
 
-void mp_emit_glue_assign_byte_code(mp_raw_code_t *rc, byte *code, uint len, int n_args, int n_locals, uint scope_flags, qstr *arg_names);
+void mp_emit_glue_assign_byte_code(mp_raw_code_t *rc, byte *code, uint len, uint n_pos_args, uint n_kwonly_args, qstr *arg_names, uint scope_flags);
 void mp_emit_glue_assign_native_code(mp_raw_code_t *rc, void *f, uint len, int n_args);
 void mp_emit_glue_assign_inline_asm_code(mp_raw_code_t *rc, void *f, uint len, int n_args);
 
diff --git a/py/emitinlinethumb.c b/py/emitinlinethumb.c
index 8b0e6af5e5..d83aa8b594 100644
--- a/py/emitinlinethumb.c
+++ b/py/emitinlinethumb.c
@@ -73,7 +73,7 @@ STATIC bool emit_inline_thumb_end_pass(emit_inline_asm_t *emit) {
 
     if (emit->pass == PASS_3) {
         void *f = asm_thumb_get_code(emit->as);
-        mp_emit_glue_assign_inline_asm_code(emit->scope->raw_code, f, asm_thumb_get_code_size(emit->as), emit->scope->num_params);
+        mp_emit_glue_assign_inline_asm_code(emit->scope->raw_code, f, asm_thumb_get_code_size(emit->as), emit->scope->num_pos_args);
     }
 
     return emit->success;
diff --git a/py/emitnative.c b/py/emitnative.c
index 79d6da67c1..1949009ebc 100644
--- a/py/emitnative.c
+++ b/py/emitnative.c
@@ -230,7 +230,7 @@ STATIC void emit_native_start_pass(emit_t *emit, pass_kind_t pass, scope_t *scop
 
     // initialise locals from parameters
 #if N_X64
-    for (int i = 0; i < scope->num_params; i++) {
+    for (int i = 0; i < scope->num_pos_args; i++) {
         if (i == 0) {
             asm_x64_mov_r64_to_r64(emit->as, REG_ARG_1, REG_LOCAL_1);
         } else if (i == 1) {
@@ -243,7 +243,7 @@ STATIC void emit_native_start_pass(emit_t *emit, pass_kind_t pass, scope_t *scop
         }
     }
 #elif N_THUMB
-    for (int i = 0; i < scope->num_params; i++) {
+    for (int i = 0; i < scope->num_pos_args; i++) {
         if (i == 0) {
             asm_thumb_mov_reg_reg(emit->as, REG_LOCAL_1, REG_ARG_1);
         } else if (i == 1) {
@@ -283,10 +283,10 @@ STATIC void emit_native_end_pass(emit_t *emit) {
     if (emit->pass == PASS_3) {
 #if N_X64
         void *f = asm_x64_get_code(emit->as);
-        mp_emit_glue_assign_native_code(emit->scope->raw_code, f, asm_x64_get_code_size(emit->as), emit->scope->num_params);
+        mp_emit_glue_assign_native_code(emit->scope->raw_code, f, asm_x64_get_code_size(emit->as), emit->scope->num_pos_args);
 #elif N_THUMB
         void *f = asm_thumb_get_code(emit->as);
-        mp_emit_glue_assign_native_code(emit->scope->raw_code, f, asm_thumb_get_code_size(emit->as), emit->scope->num_params);
+        mp_emit_glue_assign_native_code(emit->scope->raw_code, f, asm_thumb_get_code_size(emit->as), emit->scope->num_pos_args);
 #endif
     }
 }
diff --git a/py/obj.h b/py/obj.h
index 8b835b28ee..5bf5352a1e 100644
--- a/py/obj.h
+++ b/py/obj.h
@@ -356,7 +356,7 @@ mp_obj_t mp_obj_new_exception_arg1(const mp_obj_type_t *exc_type, mp_obj_t arg);
 mp_obj_t mp_obj_new_exception_args(const mp_obj_type_t *exc_type, uint n_args, const mp_obj_t *args);
 mp_obj_t mp_obj_new_exception_msg(const mp_obj_type_t *exc_type, const char *msg);
 mp_obj_t mp_obj_new_exception_msg_varg(const mp_obj_type_t *exc_type, const char *fmt, ...); // counts args by number of % symbols in fmt, excluding %%; can only handle void* sizes (ie no float/double!)
-mp_obj_t mp_obj_new_fun_bc(uint scope_flags, qstr *args, uint n_args, mp_obj_t def_args, const byte *code);
+mp_obj_t mp_obj_new_fun_bc(uint scope_flags, qstr *args, uint n_pos_args, uint n_kwonly_args, mp_obj_t def_args, const byte *code);
 mp_obj_t mp_obj_new_fun_asm(uint n_args, void *fun);
 mp_obj_t mp_obj_new_gen_wrap(mp_obj_t fun);
 mp_obj_t mp_obj_new_closure(mp_obj_t fun, uint n_closed, const mp_obj_t *closed);
diff --git a/py/objfun.c b/py/objfun.c
index 940b64a66e..a5266eb184 100644
--- a/py/objfun.c
+++ b/py/objfun.c
@@ -152,17 +152,18 @@ bool mp_obj_fun_prepare_simple_args(mp_obj_t self_in, uint n_args, uint n_kw, co
     mp_obj_fun_bc_t *self = self_in;
 
     assert(n_kw == 0);
+    assert(self->n_kwonly_args == 0);
     assert(self->takes_var_args == 0);
     assert(self->takes_kw_args == 0);
 
     mp_obj_t *extra_args = self->extra_args + self->n_def_args;
     uint n_extra_args = 0;
 
-    if (n_args > self->n_args) {
-            goto arg_error;
+    if (n_args > self->n_pos_args) {
+        goto arg_error;
     } else {
-        extra_args -= self->n_args - n_args;
-        n_extra_args += self->n_args - n_args;
+        extra_args -= self->n_pos_args - n_args;
+        n_extra_args += self->n_pos_args - n_args;
     }
     *out_args1 = args;
     *out_args1_len = n_args;
@@ -171,10 +172,15 @@ bool mp_obj_fun_prepare_simple_args(mp_obj_t self_in, uint n_args, uint n_kw, co
     return true;
 
 arg_error:
-    nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError, "function takes %d positional arguments but %d were given", self->n_args, n_args));
+    nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError, "function takes %d positional arguments but %d were given", self->n_pos_args, n_args));
 }
 
 STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_obj_t *args) {
+    // This function is pretty complicated.  It's main aim is to be efficient in speed and RAM
+    // usage for the common case of positional only args.
+    //
+    // extra_args layout: def_args, var_arg tuple, kwonly args, var_kw dict
+
     DEBUG_printf("Input n_args: %d, n_kw: %d\n", n_args, n_kw);
     DEBUG_printf("Input pos args: ");
     dump_args(args, n_args);
@@ -187,19 +193,18 @@ STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_o
     mp_obj_t *extra_args = self->extra_args + self->n_def_args;
     uint n_extra_args = 0;
 
-
     // check positional arguments
 
-    if (n_args > self->n_args) {
+    if (n_args > self->n_pos_args) {
         // given more than enough arguments
         if (!self->takes_var_args) {
             nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError,
-                "function takes %d positional arguments but %d were given", self->n_args, n_args));
+                "function takes %d positional arguments but %d were given", self->n_pos_args, n_args));
         }
         // put extra arguments in varargs tuple
-        *extra_args = mp_obj_new_tuple(n_args - self->n_args, args + self->n_args);
+        *extra_args = mp_obj_new_tuple(n_args - self->n_pos_args, args + self->n_pos_args);
         n_extra_args = 1;
-        n_args = self->n_args;
+        n_args = self->n_pos_args;
     } else {
         if (self->takes_var_args) {
             DEBUG_printf("passing empty tuple as *args\n");
@@ -209,14 +214,14 @@ STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_o
         // Apply processing and check below only if we don't have kwargs,
         // otherwise, kw handling code below has own extensive checks.
         if (n_kw == 0) {
-            if (n_args >= self->n_args - self->n_def_args) {
+            if (n_args >= self->n_pos_args - self->n_def_args) {
                 // given enough arguments, but may need to use some default arguments
-                extra_args -= self->n_args - n_args;
-                n_extra_args += self->n_args - n_args;
+                extra_args -= self->n_pos_args - n_args;
+                n_extra_args += self->n_pos_args - n_args;
             } else {
                 nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError,
                     "function takes at least %d positional arguments but %d were given",
-                    self->n_args - self->n_def_args, n_args));
+                    self->n_pos_args - self->n_def_args, n_args));
             }
         }
     }
@@ -229,13 +234,13 @@ STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_o
         // So, we have 2 choices: allocate it unconditionally at the top of function
         // (wastes stack), or use alloca which is guaranteed to dealloc on func exit.
         //mp_obj_t flat_args[self->n_args];
-        mp_obj_t *flat_args = alloca(self->n_args * sizeof(mp_obj_t));
-        for (int i = self->n_args - 1; i >= 0; i--) {
+        mp_obj_t *flat_args = alloca((self->n_pos_args + self->n_kwonly_args) * sizeof(mp_obj_t));
+        for (int i = self->n_pos_args + self->n_kwonly_args - 1; i >= 0; i--) {
             flat_args[i] = MP_OBJ_NULL;
         }
         memcpy(flat_args, args, sizeof(*args) * n_args);
         DEBUG_printf("Initial args: ");
-        dump_args(flat_args, self->n_args);
+        dump_args(flat_args, self->n_pos_args + self->n_kwonly_args);
 
         mp_obj_t dict = MP_OBJ_NULL;
         if (self->takes_kw_args) {
@@ -243,7 +248,7 @@ STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_o
         }
         for (uint i = 0; i < n_kw; i++) {
             qstr arg_name = MP_OBJ_QSTR_VALUE(kwargs[2 * i]);
-            for (uint j = 0; j < self->n_args; j++) {
+            for (uint j = 0; j < self->n_pos_args + self->n_kwonly_args; j++) {
                 if (arg_name == self->args[j]) {
                     if (flat_args[j] != MP_OBJ_NULL) {
                         nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError,
@@ -261,10 +266,10 @@ STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_o
 continue2:;
         }
         DEBUG_printf("Args with kws flattened: ");
-        dump_args(flat_args, self->n_args);
+        dump_args(flat_args, self->n_pos_args + self->n_kwonly_args);
 
-        // Now fill in defaults
-        mp_obj_t *d = &flat_args[self->n_args - 1];
+        // Now fill in defaults for positional args
+        mp_obj_t *d = &flat_args[self->n_pos_args - 1];
         mp_obj_t *s = &self->extra_args[self->n_def_args - 1];
         for (int i = self->n_def_args; i > 0; i--, d--, s--) {
             if (*d == MP_OBJ_NULL) {
@@ -272,9 +277,9 @@ continue2:;
             }
         }
         DEBUG_printf("Args after filling defaults: ");
-        dump_args(flat_args, self->n_args);
+        dump_args(flat_args, self->n_pos_args + self->n_kwonly_args);
 
-        // Now check that all mandatory args specified
+        // Check that all mandatory positional args are specified
         while (d >= flat_args) {
             if (*d-- == MP_OBJ_NULL) {
                 nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError,
@@ -282,8 +287,16 @@ continue2:;
             }
         }
 
+        // Check that all mandatory keyword args are specified
+        for (int i = 0; i < self->n_kwonly_args; i++) {
+            if (flat_args[self->n_pos_args + i] == MP_OBJ_NULL) {
+                nlr_raise(mp_obj_new_exception_msg_varg(&mp_type_TypeError,
+                    "function missing required keyword argument '%s'", qstr_str(self->args[self->n_pos_args + i])));
+            }
+        }
+
         args = flat_args;
-        n_args = self->n_args;
+        n_args = self->n_pos_args + self->n_kwonly_args;
 
         if (self->takes_kw_args) {
             extra_args[n_extra_args] = dict;
@@ -291,6 +304,10 @@ continue2:;
         }
     } else {
         // no keyword arguments given
+        if (self->n_kwonly_args != 0) {
+            nlr_raise(mp_obj_new_exception_msg(&mp_type_TypeError,
+                "function missing keyword-only argument"));
+        }
         if (self->takes_kw_args) {
             extra_args[n_extra_args] = mp_obj_new_dict(0);
             n_extra_args += 1;
@@ -320,7 +337,7 @@ const mp_obj_type_t mp_type_fun_bc = {
     .binary_op = fun_binary_op,
 };
 
-mp_obj_t mp_obj_new_fun_bc(uint scope_flags, qstr *args, uint n_args, mp_obj_t def_args_in, const byte *code) {
+mp_obj_t mp_obj_new_fun_bc(uint scope_flags, qstr *args, uint n_pos_args, uint n_kwonly_args, mp_obj_t def_args_in, const byte *code) {
     uint n_def_args = 0;
     uint n_extra_args = 0;
     mp_obj_tuple_t *def_args = def_args_in;
@@ -339,14 +356,22 @@ mp_obj_t mp_obj_new_fun_bc(uint scope_flags, qstr *args, uint n_args, mp_obj_t d
     o->base.type = &mp_type_fun_bc;
     o->globals = mp_globals_get();
     o->args = args;
-    o->n_args = n_args;
+    o->n_pos_args = n_pos_args;
+    o->n_kwonly_args = n_kwonly_args;
     o->n_def_args = n_def_args;
     o->takes_var_args = (scope_flags & MP_SCOPE_FLAG_VARARGS) != 0;
     o->takes_kw_args = (scope_flags & MP_SCOPE_FLAG_VARKEYWORDS) != 0;
     o->bytecode = code;
+    memset(o->extra_args, 0, n_extra_args * sizeof(mp_obj_t));
     if (def_args != MP_OBJ_NULL) {
         memcpy(o->extra_args, def_args->items, n_def_args * sizeof(mp_obj_t));
     }
+    if ((scope_flags & MP_SCOPE_FLAG_VARARGS) != 0) {
+        o->extra_args[n_def_args] = MP_OBJ_NULL;
+    }
+    if ((scope_flags & MP_SCOPE_FLAG_VARARGS) != 0) {
+        o->extra_args[n_extra_args - 1] = MP_OBJ_NULL;
+    }
     return o;
 }
 
diff --git a/py/objfun.h b/py/objfun.h
index 07ca623ece..8207b161a2 100644
--- a/py/objfun.h
+++ b/py/objfun.h
@@ -1,8 +1,9 @@
 typedef struct _mp_obj_fun_bc_t {
     mp_obj_base_t base;
     mp_obj_dict_t *globals; // the context within which this function was defined
-    machine_uint_t n_args : 15;         // number of arguments this function takes
-    machine_uint_t n_def_args : 15;     // number of default arguments
+    machine_uint_t n_pos_args : 16;     // number of arguments this function takes
+    machine_uint_t n_kwonly_args : 16;  // number of arguments this function takes
+    machine_uint_t n_def_args : 16;     // number of default arguments
     machine_uint_t takes_var_args : 1;  // set if this function takes variable args
     machine_uint_t takes_kw_args : 1;   // set if this function takes keyword args
     const byte *bytecode;   // bytecode for the function
diff --git a/py/scope.c b/py/scope.c
index 391ad5da87..1f92729d59 100644
--- a/py/scope.c
+++ b/py/scope.c
@@ -236,7 +236,7 @@ void scope_print_info(scope_t *s) {
     printf("\n");
     */
     printf("     flags %04x\n", s->scope_flags);
-    printf("     argcount %d\n", s->num_params);
+    printf("     argcount %d\n", s->num_pos_args);
     printf("     nlocals %d\n", s->num_locals);
     printf("     stacksize %d\n", s->stack_size);
 }
diff --git a/py/scope.h b/py/scope.h
index 6e0990ea9c..7ef81b0c83 100644
--- a/py/scope.h
+++ b/py/scope.h
@@ -8,7 +8,8 @@ enum {
 
 enum {
     ID_FLAG_IS_PARAM = 0x01,
-    ID_FLAG_IS_DELETED = 0x02,
+    ID_FLAG_IS_STAR_PARAM = 0x02,
+    ID_FLAG_IS_DBL_STAR_PARAM = 0x04,
 };
 
 typedef struct _id_info_t {
@@ -32,7 +33,8 @@ typedef struct _scope_t {
     mp_raw_code_t *raw_code;
     uint8_t scope_flags;  // see runtime0.h
     uint8_t emit_options; // see compile.h
-    uint16_t num_params;
+    uint16_t num_pos_args;
+    uint16_t num_kwonly_args;
     uint16_t num_locals;
     uint16_t stack_size;     // maximum size of the locals stack
     uint16_t exc_stack_size; // maximum size of the exception stack
diff --git a/tests/basics/fun-kwonly.py b/tests/basics/fun-kwonly.py
new file mode 100644
index 0000000000..8bda68d95a
--- /dev/null
+++ b/tests/basics/fun-kwonly.py
@@ -0,0 +1,68 @@
+# to test keyword-only arguments
+
+# simplest case
+def f(*, a):
+    print(a)
+
+f(a=1)
+
+# with 2 keyword-only args
+def f(*, a, b):
+    print(a, b)
+
+f(a=1, b=2)
+f(b=1, a=2)
+
+# positional followed by bare star
+def f(a, *, b, c):
+    print(a, b, c)
+
+f(1, b=3, c=4)
+f(1, c=3, b=4)
+f(1, **{'b':'3', 'c':4})
+
+try:
+    f(1)
+except TypeError:
+    print("TypeError")
+
+try:
+    f(1, b=2)
+except TypeError:
+    print("TypeError")
+
+try:
+    f(1, c=2)
+except TypeError:
+    print("TypeError")
+
+# with **kw
+def f(a, *, b, **kw):
+    print(a, b, kw)
+
+f(1, b=2)
+f(1, b=2, c=3)
+
+## with a default value; not currently working
+#def g(a, *, b=2, c):
+#    print(a, b, c)
+#
+#g(1, c=3)
+#g(1, b=3, c=4)
+#g(1, **{'c':3})
+#g(1, **{'b':'3', 'c':4})
+
+# with named star
+def f(*a, b, c):
+    print(a, b, c)
+
+f(b=1, c=2)
+f(c=1, b=2)
+
+# with positional and named star
+def f(a, *b, c):
+    print(a, b, c)
+
+f(1, c=2)
+f(1, 2, c=3)
+f(a=1, c=3)