py: Simplify fastn in VM; reduce size of unique code struct.

We still have FAST_[0,1,2] byte codes, but they now just access the
fastn array (before they had special local variables).  It's now
simpler, a bit faster, and uses a bit less stack space (on STM at least,
which is most important).

The only reason now to keep FAST_[0,1,2] byte codes is for compressed
byte code size.
This commit is contained in:
Damien George 2014-01-29 20:30:52 +00:00
parent 01fa4a9164
commit d0691ccaec
5 changed files with 38 additions and 64 deletions

View File

@ -223,7 +223,7 @@ mp_obj_t mp_obj_new_range(int start, int stop, int step);
mp_obj_t mp_obj_new_range_iterator(int cur, int stop, int step);
mp_obj_t mp_obj_new_fun_bc(int n_args, uint n_state, const byte *code);
mp_obj_t mp_obj_new_fun_asm(uint n_args, void *fun);
mp_obj_t mp_obj_new_gen_wrap(uint n_locals, uint n_stack, mp_obj_t fun);
mp_obj_t mp_obj_new_gen_wrap(mp_obj_t fun);
mp_obj_t mp_obj_new_gen_instance(const byte *bytecode, uint n_state, int n_args, const mp_obj_t *args);
mp_obj_t mp_obj_new_closure(mp_obj_t fun, mp_obj_t closure_tuple);
mp_obj_t mp_obj_new_tuple(uint n, const mp_obj_t *items);

View File

@ -16,7 +16,6 @@
typedef struct _mp_obj_gen_wrap_t {
mp_obj_base_t base;
uint n_state;
mp_obj_t *fun;
} mp_obj_gen_wrap_t;
@ -35,7 +34,7 @@ mp_obj_t gen_wrap_call(mp_obj_t self_in, uint n_args, uint n_kw, const mp_obj_t
nlr_jump(mp_obj_new_exception_msg(MP_QSTR_TypeError, "function does not take keyword arguments"));
}
return mp_obj_new_gen_instance(bc_code, self->n_state, n_args, args);
return mp_obj_new_gen_instance(bc_code, bc_n_state, n_args, args);
}
const mp_obj_type_t gen_wrap_type = {
@ -44,11 +43,9 @@ const mp_obj_type_t gen_wrap_type = {
.call = gen_wrap_call,
};
mp_obj_t mp_obj_new_gen_wrap(uint n_locals, uint n_stack, mp_obj_t fun) {
mp_obj_t mp_obj_new_gen_wrap(mp_obj_t fun) {
mp_obj_gen_wrap_t *o = m_new_obj(mp_obj_gen_wrap_t);
o->base.type = &gen_wrap_type;
// we have at least 3 locals so the bc can write back fast[0,1,2] safely; should improve how this is done
o->n_state = (n_locals < 3 ? 3 : n_locals) + n_stack;
o->fun = fun;
return o;
}

View File

@ -44,11 +44,14 @@ typedef enum {
} mp_code_kind_t;
typedef struct _mp_code_t {
mp_code_kind_t kind;
int n_args;
int n_locals;
int n_stack;
bool is_generator;
struct {
mp_code_kind_t kind : 8;
bool is_generator : 1;
};
struct {
uint n_args : 16;
uint n_state : 16;
};
union {
struct {
byte *code;
@ -63,7 +66,7 @@ typedef struct _mp_code_t {
};
} mp_code_t;
static int next_unique_code_id;
static uint next_unique_code_id;
static machine_uint_t unique_codes_alloc = 0;
static mp_code_t *unique_codes = NULL;
@ -187,30 +190,30 @@ void rt_deinit(void) {
#endif
}
int rt_get_unique_code_id(void) {
uint rt_get_unique_code_id(void) {
return next_unique_code_id++;
}
static void alloc_unique_codes(void) {
if (next_unique_code_id > unique_codes_alloc) {
DEBUG_printf("allocate more unique codes: " UINT_FMT " -> %u\n", unique_codes_alloc, next_unique_code_id);
// increase size of unique_codes table
unique_codes = m_renew(mp_code_t, unique_codes, unique_codes_alloc, next_unique_code_id);
for (int i = unique_codes_alloc; i < next_unique_code_id; i++) {
for (uint i = unique_codes_alloc; i < next_unique_code_id; i++) {
unique_codes[i].kind = MP_CODE_NONE;
}
unique_codes_alloc = next_unique_code_id;
}
}
void rt_assign_byte_code(int unique_code_id, byte *code, uint len, int n_args, int n_locals, int n_stack, bool is_generator) {
void rt_assign_byte_code(uint unique_code_id, byte *code, uint len, int n_args, int n_locals, int n_stack, bool is_generator) {
alloc_unique_codes();
assert(1 <= unique_code_id && unique_code_id < next_unique_code_id && unique_codes[unique_code_id].kind == MP_CODE_NONE);
unique_codes[unique_code_id].kind = MP_CODE_BYTE;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_locals = n_locals;
unique_codes[unique_code_id].n_stack = n_stack;
unique_codes[unique_code_id].is_generator = is_generator;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_state = n_locals + n_stack;
unique_codes[unique_code_id].u_byte.code = code;
unique_codes[unique_code_id].u_byte.len = len;
@ -238,15 +241,14 @@ void rt_assign_byte_code(int unique_code_id, byte *code, uint len, int n_args, i
#endif
}
void rt_assign_native_code(int unique_code_id, void *fun, uint len, int n_args) {
void rt_assign_native_code(uint unique_code_id, void *fun, uint len, int n_args) {
alloc_unique_codes();
assert(1 <= unique_code_id && unique_code_id < next_unique_code_id && unique_codes[unique_code_id].kind == MP_CODE_NONE);
unique_codes[unique_code_id].kind = MP_CODE_NATIVE;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_locals = 0;
unique_codes[unique_code_id].n_stack = 0;
unique_codes[unique_code_id].is_generator = false;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_state = 0;
unique_codes[unique_code_id].u_native.fun = fun;
//printf("native code: %d bytes\n", len);
@ -271,15 +273,14 @@ void rt_assign_native_code(int unique_code_id, void *fun, uint len, int n_args)
#endif
}
void rt_assign_inline_asm_code(int unique_code_id, void *fun, uint len, int n_args) {
void rt_assign_inline_asm_code(uint unique_code_id, void *fun, uint len, int n_args) {
alloc_unique_codes();
assert(1 <= unique_code_id && unique_code_id < next_unique_code_id && unique_codes[unique_code_id].kind == MP_CODE_NONE);
unique_codes[unique_code_id].kind = MP_CODE_INLINE_ASM;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_locals = 0;
unique_codes[unique_code_id].n_stack = 0;
unique_codes[unique_code_id].is_generator = false;
unique_codes[unique_code_id].n_args = n_args;
unique_codes[unique_code_id].n_state = 0;
unique_codes[unique_code_id].u_inline_asm.fun = fun;
#ifdef DEBUG_PRINT
@ -678,7 +679,7 @@ mp_obj_t rt_make_function_from_id(int unique_code_id) {
mp_obj_t fun;
switch (c->kind) {
case MP_CODE_BYTE:
fun = mp_obj_new_fun_bc(c->n_args, c->n_locals + c->n_stack, c->u_byte.code);
fun = mp_obj_new_fun_bc(c->n_args, c->n_state, c->u_byte.code);
break;
case MP_CODE_NATIVE:
fun = rt_make_function_n(c->n_args, c->u_native.fun);
@ -693,7 +694,7 @@ mp_obj_t rt_make_function_from_id(int unique_code_id) {
// check for generator functions and if so wrap in generator object
if (c->is_generator) {
fun = mp_obj_new_gen_wrap(c->n_locals, c->n_stack, fun);
fun = mp_obj_new_gen_wrap(fun);
}
return fun;

View File

@ -78,7 +78,7 @@ extern void *const rt_fun_table[RT_F_NUMBER_OF];
void rt_init(void);
void rt_deinit(void);
int rt_get_unique_code_id(void);
void rt_assign_byte_code(int unique_code_id, byte *code, uint len, int n_args, int n_locals, int n_stack, bool is_generator);
void rt_assign_native_code(int unique_code_id, void *f, uint len, int n_args);
void rt_assign_inline_asm_code(int unique_code_id, void *f, uint len, int n_args);
uint rt_get_unique_code_id(void);
void rt_assign_byte_code(uint unique_code_id, byte *code, uint len, int n_args, int n_locals, int n_stack, bool is_generator);
void rt_assign_native_code(uint unique_code_id, void *f, uint len, int n_args);
void rt_assign_inline_asm_code(uint unique_code_id, void *f, uint len, int n_args);

40
py/vm.c
View File

@ -80,7 +80,6 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
machine_uint_t unum;
qstr qst;
mp_obj_t obj1, obj2;
mp_obj_t fast0 = fastn[0], fast1 = fastn[-1], fast2 = fastn[-2];
nlr_buf_t nlr;
volatile machine_uint_t currently_in_except_block = 0; // 0 or 1, to detect nested exceptions
@ -88,8 +87,6 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
machine_uint_t *volatile exc_sp = &exc_stack[0] - 1; // stack grows up, exc_sp points to top of stack
const byte *volatile save_ip = ip; // this is so we can access ip in the exception handler without making ip volatile (which means the compiler can't keep it in a register in the main loop)
// TODO if an exception occurs, do fast[0,1,2] become invalid??
// outer exception handling loop
for (;;) {
if (nlr_push(&nlr) == 0) {
@ -146,15 +143,15 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
break;
case MP_BC_LOAD_FAST_0:
PUSH(fast0);
PUSH(fastn[0]);
break;
case MP_BC_LOAD_FAST_1:
PUSH(fast1);
PUSH(fastn[-1]);
break;
case MP_BC_LOAD_FAST_2:
PUSH(fast2);
PUSH(fastn[-2]);
break;
case MP_BC_LOAD_FAST_N:
@ -164,16 +161,7 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
case MP_BC_LOAD_DEREF:
DECODE_UINT;
if (unum == 0) {
obj1 = fast0;
} else if (unum == 1) {
obj1 = fast1;
} else if (unum == 2) {
obj1 = fast2;
} else {
obj1 = fastn[-unum];
}
PUSH(rt_get_cell(obj1));
PUSH(rt_get_cell(fastn[-unum]));
break;
case MP_BC_LOAD_NAME:
@ -202,15 +190,15 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
break;
case MP_BC_STORE_FAST_0:
fast0 = POP();
fastn[0] = POP();
break;
case MP_BC_STORE_FAST_1:
fast1 = POP();
fastn[-1] = POP();
break;
case MP_BC_STORE_FAST_2:
fast2 = POP();
fastn[-2] = POP();
break;
case MP_BC_STORE_FAST_N:
@ -220,16 +208,7 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
case MP_BC_STORE_DEREF:
DECODE_UINT;
if (unum == 0) {
obj1 = fast0;
} else if (unum == 1) {
obj1 = fast1;
} else if (unum == 2) {
obj1 = fast2;
} else {
obj1 = fastn[-unum];
}
rt_set_cell(obj1, POP());
rt_set_cell(fastn[-unum], POP());
break;
case MP_BC_STORE_NAME:
@ -511,9 +490,6 @@ bool mp_execute_byte_code_2(const byte *code_info, const byte **ip_in_out, mp_ob
case MP_BC_YIELD_VALUE:
nlr_pop();
*ip_in_out = ip;
fastn[0] = fast0;
fastn[-1] = fast1;
fastn[-2] = fast2;
*sp_in_out = sp;
return true;