2014-05-03 23:27:38 +01:00
|
|
|
/*
|
2017-06-30 08:22:17 +01:00
|
|
|
* This file is part of the MicroPython project, http://micropython.org/
|
2014-05-03 23:27:38 +01:00
|
|
|
*
|
|
|
|
* The MIT License (MIT)
|
|
|
|
*
|
|
|
|
* Copyright (c) 2013, 2014 Damien P. George
|
2014-05-13 06:44:45 +01:00
|
|
|
* Copyright (c) 2014 Paul Sokolovsky
|
2014-05-03 23:27:38 +01:00
|
|
|
*
|
|
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
|
|
* of this software and associated documentation files (the "Software"), to deal
|
|
|
|
* in the Software without restriction, including without limitation the rights
|
|
|
|
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
|
|
* copies of the Software, and to permit persons to whom the Software is
|
|
|
|
* furnished to do so, subject to the following conditions:
|
|
|
|
*
|
|
|
|
* The above copyright notice and this permission notice shall be included in
|
|
|
|
* all copies or substantial portions of the Software.
|
|
|
|
*
|
|
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
|
|
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
|
|
|
* THE SOFTWARE.
|
|
|
|
*/
|
|
|
|
|
2013-12-21 18:17:45 +00:00
|
|
|
#include <string.h>
|
|
|
|
#include <assert.h>
|
|
|
|
|
2015-01-01 20:27:54 +00:00
|
|
|
#include "py/objtuple.h"
|
|
|
|
#include "py/objfun.h"
|
|
|
|
#include "py/runtime.h"
|
|
|
|
#include "py/bc.h"
|
|
|
|
#include "py/stackctrl.h"
|
2013-12-21 18:17:45 +00:00
|
|
|
|
2017-07-24 17:55:14 +01:00
|
|
|
#if MICROPY_DEBUG_VERBOSE // print debugging info
|
2014-02-16 16:30:49 +00:00
|
|
|
#define DEBUG_PRINT (1)
|
|
|
|
#else // don't print debugging info
|
2014-11-05 21:16:41 +00:00
|
|
|
#define DEBUG_PRINT (0)
|
2014-02-26 22:40:35 +00:00
|
|
|
#define DEBUG_printf(...) (void)0
|
2014-02-16 16:30:49 +00:00
|
|
|
#endif
|
|
|
|
|
2015-01-11 21:07:15 +00:00
|
|
|
// Note: the "name" entry in mp_obj_type_t for a function type must be
|
|
|
|
// MP_QSTR_function because it is used to determine if an object is of generic
|
|
|
|
// function type.
|
2014-04-05 11:50:06 +01:00
|
|
|
|
2014-08-24 16:28:17 +01:00
|
|
|
/******************************************************************************/
|
|
|
|
/* builtin functions */
|
|
|
|
|
2016-10-18 01:49:27 +01:00
|
|
|
STATIC mp_obj_t fun_builtin_0_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
|
|
|
(void)args;
|
2019-01-30 07:49:52 +00:00
|
|
|
assert(mp_obj_is_type(self_in, &mp_type_fun_builtin_0));
|
2016-10-18 01:49:27 +01:00
|
|
|
mp_obj_fun_builtin_fixed_t *self = MP_OBJ_TO_PTR(self_in);
|
|
|
|
mp_arg_check_num(n_args, n_kw, 0, 0, false);
|
|
|
|
return self->fun._0();
|
|
|
|
}
|
|
|
|
|
|
|
|
const mp_obj_type_t mp_type_fun_builtin_0 = {
|
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF | MP_TYPE_FLAG_BUILTIN_FUN,
|
2016-10-18 01:49:27 +01:00
|
|
|
.name = MP_QSTR_function,
|
|
|
|
.call = fun_builtin_0_call,
|
|
|
|
.unary_op = mp_generic_unary_op,
|
|
|
|
};
|
|
|
|
|
|
|
|
STATIC mp_obj_t fun_builtin_1_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2019-01-30 07:49:52 +00:00
|
|
|
assert(mp_obj_is_type(self_in, &mp_type_fun_builtin_1));
|
2016-10-18 01:49:27 +01:00
|
|
|
mp_obj_fun_builtin_fixed_t *self = MP_OBJ_TO_PTR(self_in);
|
|
|
|
mp_arg_check_num(n_args, n_kw, 1, 1, false);
|
|
|
|
return self->fun._1(args[0]);
|
|
|
|
}
|
|
|
|
|
|
|
|
const mp_obj_type_t mp_type_fun_builtin_1 = {
|
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF | MP_TYPE_FLAG_BUILTIN_FUN,
|
2016-10-18 01:49:27 +01:00
|
|
|
.name = MP_QSTR_function,
|
|
|
|
.call = fun_builtin_1_call,
|
|
|
|
.unary_op = mp_generic_unary_op,
|
|
|
|
};
|
|
|
|
|
|
|
|
STATIC mp_obj_t fun_builtin_2_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2019-01-30 07:49:52 +00:00
|
|
|
assert(mp_obj_is_type(self_in, &mp_type_fun_builtin_2));
|
2016-10-18 01:49:27 +01:00
|
|
|
mp_obj_fun_builtin_fixed_t *self = MP_OBJ_TO_PTR(self_in);
|
|
|
|
mp_arg_check_num(n_args, n_kw, 2, 2, false);
|
|
|
|
return self->fun._2(args[0], args[1]);
|
|
|
|
}
|
|
|
|
|
|
|
|
const mp_obj_type_t mp_type_fun_builtin_2 = {
|
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF | MP_TYPE_FLAG_BUILTIN_FUN,
|
2016-10-18 01:49:27 +01:00
|
|
|
.name = MP_QSTR_function,
|
|
|
|
.call = fun_builtin_2_call,
|
|
|
|
.unary_op = mp_generic_unary_op,
|
|
|
|
};
|
|
|
|
|
|
|
|
STATIC mp_obj_t fun_builtin_3_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2019-01-30 07:49:52 +00:00
|
|
|
assert(mp_obj_is_type(self_in, &mp_type_fun_builtin_3));
|
2016-10-18 01:49:27 +01:00
|
|
|
mp_obj_fun_builtin_fixed_t *self = MP_OBJ_TO_PTR(self_in);
|
|
|
|
mp_arg_check_num(n_args, n_kw, 3, 3, false);
|
|
|
|
return self->fun._3(args[0], args[1], args[2]);
|
|
|
|
}
|
|
|
|
|
|
|
|
const mp_obj_type_t mp_type_fun_builtin_3 = {
|
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF | MP_TYPE_FLAG_BUILTIN_FUN,
|
2016-10-18 01:49:27 +01:00
|
|
|
.name = MP_QSTR_function,
|
|
|
|
.call = fun_builtin_3_call,
|
|
|
|
.unary_op = mp_generic_unary_op,
|
|
|
|
};
|
2014-08-24 16:28:17 +01:00
|
|
|
|
2016-10-18 01:49:27 +01:00
|
|
|
STATIC mp_obj_t fun_builtin_var_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2019-01-30 07:49:52 +00:00
|
|
|
assert(mp_obj_is_type(self_in, &mp_type_fun_builtin_var));
|
2016-10-18 01:49:27 +01:00
|
|
|
mp_obj_fun_builtin_var_t *self = MP_OBJ_TO_PTR(self_in);
|
2014-01-18 14:10:48 +00:00
|
|
|
|
2014-01-13 19:55:18 +00:00
|
|
|
// check number of arguments
|
2018-09-14 04:39:17 +01:00
|
|
|
mp_arg_check_num_sig(n_args, n_kw, self->sig);
|
2014-01-18 14:10:48 +00:00
|
|
|
|
2018-09-14 04:39:17 +01:00
|
|
|
if (self->sig & 1) {
|
2014-01-18 14:10:48 +00:00
|
|
|
// function allows keywords
|
|
|
|
|
2014-02-08 18:53:41 +00:00
|
|
|
// we create a map directly from the given args array
|
|
|
|
mp_map_t kw_args;
|
|
|
|
mp_map_init_fixed_table(&kw_args, n_kw, args + n_args);
|
2014-01-18 14:10:48 +00:00
|
|
|
|
2016-01-03 11:53:44 +00:00
|
|
|
return self->fun.kw(n_args, args, &kw_args);
|
2014-01-18 14:10:48 +00:00
|
|
|
|
2013-12-21 18:17:45 +00:00
|
|
|
} else {
|
2014-01-18 14:10:48 +00:00
|
|
|
// function takes a variable number of arguments, but no keywords
|
2013-12-21 18:17:45 +00:00
|
|
|
|
2016-01-03 11:53:44 +00:00
|
|
|
return self->fun.var(n_args, args);
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-10-18 01:49:27 +01:00
|
|
|
const mp_obj_type_t mp_type_fun_builtin_var = {
|
2014-02-15 16:10:44 +00:00
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF | MP_TYPE_FLAG_BUILTIN_FUN,
|
2014-02-15 11:34:50 +00:00
|
|
|
.name = MP_QSTR_function,
|
2016-10-18 01:49:27 +01:00
|
|
|
.call = fun_builtin_var_call,
|
2015-05-11 13:25:19 +01:00
|
|
|
.unary_op = mp_generic_unary_op,
|
2013-12-21 18:17:45 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
/******************************************************************************/
|
|
|
|
/* byte code functions */
|
|
|
|
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
STATIC qstr mp_obj_code_get_name(const mp_obj_fun_bc_t *fun, const byte *code_info) {
|
2019-09-25 06:45:47 +01:00
|
|
|
MP_BC_PRELUDE_SIZE_DECODE(code_info);
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_uint_t name = mp_decode_uint_value(code_info);
|
|
|
|
#if MICROPY_EMIT_BYTECODE_USES_QSTR_TABLE
|
|
|
|
name = fun->context->constants.qstr_table[name];
|
2015-11-02 17:27:18 +00:00
|
|
|
#endif
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
return name;
|
2014-04-30 23:32:58 +01:00
|
|
|
}
|
|
|
|
|
2015-04-06 22:38:53 +01:00
|
|
|
#if MICROPY_EMIT_NATIVE
|
|
|
|
STATIC const mp_obj_type_t mp_type_fun_native;
|
|
|
|
#endif
|
|
|
|
|
2015-02-14 17:44:31 +00:00
|
|
|
qstr mp_obj_fun_get_name(mp_const_obj_t fun_in) {
|
2015-11-27 17:01:44 +00:00
|
|
|
const mp_obj_fun_bc_t *fun = MP_OBJ_TO_PTR(fun_in);
|
2015-04-06 22:38:53 +01:00
|
|
|
#if MICROPY_EMIT_NATIVE
|
2018-10-01 04:07:04 +01:00
|
|
|
if (fun->base.type == &mp_type_fun_native || fun->base.type == &mp_type_native_gen_wrap) {
|
2015-04-06 22:38:53 +01:00
|
|
|
// TODO native functions don't have name stored
|
|
|
|
return MP_QSTR_;
|
|
|
|
}
|
|
|
|
#endif
|
2015-03-18 17:47:47 +00:00
|
|
|
|
|
|
|
const byte *bc = fun->bytecode;
|
2019-09-16 13:12:59 +01:00
|
|
|
MP_BC_PRELUDE_SIG_DECODE(bc);
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
return mp_obj_code_get_name(fun, bc);
|
2014-05-01 20:20:07 +01:00
|
|
|
}
|
|
|
|
|
2014-04-30 23:32:58 +01:00
|
|
|
#if MICROPY_CPYTHON_COMPAT
|
2015-04-09 23:56:15 +01:00
|
|
|
STATIC void fun_bc_print(const mp_print_t *print, mp_obj_t o_in, mp_print_kind_t kind) {
|
2015-01-20 12:47:20 +00:00
|
|
|
(void)kind;
|
2015-11-27 17:01:44 +00:00
|
|
|
mp_obj_fun_bc_t *o = MP_OBJ_TO_PTR(o_in);
|
|
|
|
mp_printf(print, "<function %q at 0x%p>", mp_obj_fun_get_name(o_in), o);
|
2014-04-30 23:32:58 +01:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2014-02-16 16:30:49 +00:00
|
|
|
#if DEBUG_PRINT
|
2017-02-16 05:34:53 +00:00
|
|
|
STATIC void dump_args(const mp_obj_t *a, size_t sz) {
|
2014-02-16 16:30:49 +00:00
|
|
|
DEBUG_printf("%p: ", a);
|
2017-02-16 05:34:53 +00:00
|
|
|
for (size_t i = 0; i < sz; i++) {
|
2014-02-16 16:30:49 +00:00
|
|
|
DEBUG_printf("%p ", a[i]);
|
|
|
|
}
|
|
|
|
DEBUG_printf("\n");
|
|
|
|
}
|
2014-03-31 15:59:25 +01:00
|
|
|
#else
|
|
|
|
#define dump_args(...) (void)0
|
|
|
|
#endif
|
2014-02-16 16:30:49 +00:00
|
|
|
|
2014-06-07 14:16:08 +01:00
|
|
|
// With this macro you can tune the maximum number of function state bytes
|
|
|
|
// that will be allocated on the stack. Any function that needs more
|
2015-06-08 22:07:27 +01:00
|
|
|
// than this will try to use the heap, with fallback to stack allocation.
|
2020-03-09 23:09:49 +00:00
|
|
|
#define VM_MAX_STATE_ON_STACK (sizeof(mp_uint_t) * 11)
|
2014-06-07 14:16:08 +01:00
|
|
|
|
2017-12-09 07:19:34 +00:00
|
|
|
#define DECODE_CODESTATE_SIZE(bytecode, n_state_out_var, state_size_out_var) \
|
|
|
|
{ \
|
2019-09-16 13:12:59 +01:00
|
|
|
const uint8_t *ip = bytecode; \
|
|
|
|
size_t n_exc_stack, scope_flags, n_pos_args, n_kwonly_args, n_def_args; \
|
|
|
|
MP_BC_PRELUDE_SIG_DECODE_INTO(ip, n_state_out_var, n_exc_stack, scope_flags, n_pos_args, n_kwonly_args, n_def_args); \
|
2021-06-24 02:45:13 +01:00
|
|
|
(void)scope_flags; (void)n_pos_args; (void)n_kwonly_args; (void)n_def_args; \
|
|
|
|
\
|
2017-12-09 07:19:34 +00:00
|
|
|
/* state size in bytes */ \
|
2020-06-13 19:05:48 +01:00
|
|
|
state_size_out_var = n_state_out_var * sizeof(mp_obj_t) \
|
|
|
|
+ n_exc_stack * sizeof(mp_exc_stack_t); \
|
2017-12-09 07:19:34 +00:00
|
|
|
}
|
|
|
|
|
2019-09-24 06:57:08 +01:00
|
|
|
#define INIT_CODESTATE(code_state, _fun_bc, _n_state, n_args, n_kw, args) \
|
2017-12-09 10:45:28 +00:00
|
|
|
code_state->fun_bc = _fun_bc; \
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
code_state->ip = _fun_bc->bytecode; \
|
2019-09-24 06:57:08 +01:00
|
|
|
code_state->n_state = _n_state; \
|
2017-12-09 10:45:28 +00:00
|
|
|
mp_setup_code_state(code_state, n_args, n_kw, args); \
|
|
|
|
code_state->old_globals = mp_globals_get();
|
|
|
|
|
2015-04-01 23:31:30 +01:00
|
|
|
#if MICROPY_STACKLESS
|
2016-08-27 14:21:00 +01:00
|
|
|
mp_code_state_t *mp_obj_fun_bc_prepare_codestate(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2015-03-27 23:14:44 +00:00
|
|
|
MP_STACK_CHECK();
|
2015-12-17 12:32:41 +00:00
|
|
|
mp_obj_fun_bc_t *self = MP_OBJ_TO_PTR(self_in);
|
2015-03-27 23:14:44 +00:00
|
|
|
|
2017-12-09 07:19:34 +00:00
|
|
|
size_t n_state, state_size;
|
|
|
|
DECODE_CODESTATE_SIZE(self->bytecode, n_state, state_size);
|
2015-03-27 23:14:44 +00:00
|
|
|
|
2016-08-27 14:21:00 +01:00
|
|
|
mp_code_state_t *code_state;
|
2017-11-26 12:37:19 +00:00
|
|
|
#if MICROPY_ENABLE_PYSTACK
|
|
|
|
code_state = mp_pystack_alloc(sizeof(mp_code_state_t) + state_size);
|
|
|
|
#else
|
2017-12-09 08:57:40 +00:00
|
|
|
// If we use m_new_obj_var(), then on no memory, MemoryError will be
|
|
|
|
// raised. But this is not correct exception for a function call,
|
|
|
|
// RuntimeError should be raised instead. So, we use m_new_obj_var_maybe(),
|
|
|
|
// return NULL, then vm.c takes the needed action (either raise
|
|
|
|
// RuntimeError or fallback to stack allocation).
|
2016-08-27 14:21:00 +01:00
|
|
|
code_state = m_new_obj_var_maybe(mp_code_state_t, byte, state_size);
|
2015-03-27 23:14:44 +00:00
|
|
|
if (!code_state) {
|
|
|
|
return NULL;
|
|
|
|
}
|
2017-11-26 12:37:19 +00:00
|
|
|
#endif
|
2015-03-27 23:14:44 +00:00
|
|
|
|
2019-09-24 06:57:08 +01:00
|
|
|
INIT_CODESTATE(code_state, self, n_state, n_args, n_kw, args);
|
2015-03-27 23:14:44 +00:00
|
|
|
|
|
|
|
// execute the byte code with the correct globals context
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_globals_set(self->context->module.globals);
|
2015-03-27 23:14:44 +00:00
|
|
|
|
|
|
|
return code_state;
|
|
|
|
}
|
2015-04-01 23:31:30 +01:00
|
|
|
#endif
|
2015-03-27 23:14:44 +00:00
|
|
|
|
2016-01-03 09:59:18 +00:00
|
|
|
STATIC mp_obj_t fun_bc_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2014-07-01 00:13:42 +01:00
|
|
|
MP_STACK_CHECK();
|
2014-06-11 15:56:43 +01:00
|
|
|
|
2014-09-23 10:59:05 +01:00
|
|
|
DEBUG_printf("Input n_args: " UINT_FMT ", n_kw: " UINT_FMT "\n", n_args, n_kw);
|
2014-06-11 15:56:43 +01:00
|
|
|
DEBUG_printf("Input pos args: ");
|
|
|
|
dump_args(args, n_args);
|
|
|
|
DEBUG_printf("Input kw args: ");
|
|
|
|
dump_args(args + n_args, n_kw * 2);
|
2018-08-02 05:17:24 +01:00
|
|
|
|
2015-11-27 17:01:44 +00:00
|
|
|
mp_obj_fun_bc_t *self = MP_OBJ_TO_PTR(self_in);
|
2014-06-11 15:56:43 +01:00
|
|
|
|
2017-12-09 07:19:34 +00:00
|
|
|
size_t n_state, state_size;
|
|
|
|
DECODE_CODESTATE_SIZE(self->bytecode, n_state, state_size);
|
2014-06-11 15:56:43 +01:00
|
|
|
|
|
|
|
// allocate state for locals and stack
|
2016-08-27 14:21:00 +01:00
|
|
|
mp_code_state_t *code_state = NULL;
|
2017-11-26 12:37:19 +00:00
|
|
|
#if MICROPY_ENABLE_PYSTACK
|
|
|
|
code_state = mp_pystack_alloc(sizeof(mp_code_state_t) + state_size);
|
|
|
|
#else
|
2014-06-11 15:56:43 +01:00
|
|
|
if (state_size > VM_MAX_STATE_ON_STACK) {
|
2016-08-27 14:21:00 +01:00
|
|
|
code_state = m_new_obj_var_maybe(mp_code_state_t, byte, state_size);
|
2019-01-04 06:09:41 +00:00
|
|
|
#if MICROPY_DEBUG_VM_STACK_OVERFLOW
|
|
|
|
if (code_state != NULL) {
|
|
|
|
memset(code_state->state, 0, state_size);
|
|
|
|
}
|
|
|
|
#endif
|
2015-06-08 22:07:27 +01:00
|
|
|
}
|
|
|
|
if (code_state == NULL) {
|
2016-08-27 14:21:00 +01:00
|
|
|
code_state = alloca(sizeof(mp_code_state_t) + state_size);
|
2019-01-04 06:09:41 +00:00
|
|
|
#if MICROPY_DEBUG_VM_STACK_OVERFLOW
|
|
|
|
memset(code_state->state, 0, state_size);
|
|
|
|
#endif
|
2015-06-08 22:07:27 +01:00
|
|
|
state_size = 0; // indicate that we allocated using alloca
|
2014-06-11 15:56:43 +01:00
|
|
|
}
|
2017-11-26 12:37:19 +00:00
|
|
|
#endif
|
2014-06-11 15:56:43 +01:00
|
|
|
|
2019-09-24 06:57:08 +01:00
|
|
|
INIT_CODESTATE(code_state, self, n_state, n_args, n_kw, args);
|
2014-06-08 00:37:40 +01:00
|
|
|
|
|
|
|
// execute the byte code with the correct globals context
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_globals_set(self->context->module.globals);
|
2014-06-07 14:16:08 +01:00
|
|
|
mp_vm_return_kind_t vm_return_kind = mp_execute_bytecode(code_state, MP_OBJ_NULL);
|
2015-02-15 09:12:58 +00:00
|
|
|
mp_globals_set(code_state->old_globals);
|
2014-06-07 14:16:08 +01:00
|
|
|
|
2019-01-04 06:09:41 +00:00
|
|
|
#if MICROPY_DEBUG_VM_STACK_OVERFLOW
|
2014-06-07 14:16:08 +01:00
|
|
|
if (vm_return_kind == MP_VM_RETURN_NORMAL) {
|
|
|
|
if (code_state->sp < code_state->state) {
|
2019-01-04 06:09:41 +00:00
|
|
|
mp_printf(MICROPY_DEBUG_PRINTER, "VM stack underflow: " INT_FMT "\n", code_state->sp - code_state->state);
|
2014-06-07 14:16:08 +01:00
|
|
|
assert(0);
|
|
|
|
}
|
|
|
|
}
|
2019-09-16 13:12:59 +01:00
|
|
|
const byte *bytecode_ptr = self->bytecode;
|
|
|
|
size_t n_state_unused, n_exc_stack_unused, scope_flags_unused;
|
|
|
|
size_t n_pos_args, n_kwonly_args, n_def_args_unused;
|
|
|
|
MP_BC_PRELUDE_SIG_DECODE_INTO(bytecode_ptr, n_state_unused, n_exc_stack_unused,
|
|
|
|
scope_flags_unused, n_pos_args, n_kwonly_args, n_def_args_unused);
|
2019-01-04 06:09:41 +00:00
|
|
|
// We can't check the case when an exception is returned in state[0]
|
2014-06-07 14:16:08 +01:00
|
|
|
// and there are no arguments, because in this case our detection slot may have
|
|
|
|
// been overwritten by the returned exception (which is allowed).
|
2019-01-04 06:09:41 +00:00
|
|
|
if (!(vm_return_kind == MP_VM_RETURN_EXCEPTION && n_pos_args + n_kwonly_args == 0)) {
|
2014-06-07 14:16:08 +01:00
|
|
|
// Just check to see that we have at least 1 null object left in the state.
|
|
|
|
bool overflow = true;
|
2019-01-04 06:09:41 +00:00
|
|
|
for (size_t i = 0; i < n_state - n_pos_args - n_kwonly_args; ++i) {
|
2014-06-07 14:16:08 +01:00
|
|
|
if (code_state->state[i] == MP_OBJ_NULL) {
|
|
|
|
overflow = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (overflow) {
|
2019-01-04 06:09:41 +00:00
|
|
|
mp_printf(MICROPY_DEBUG_PRINTER, "VM stack overflow state=%p n_state+1=" UINT_FMT "\n", code_state->state, n_state);
|
2014-06-07 14:16:08 +01:00
|
|
|
assert(0);
|
|
|
|
}
|
|
|
|
}
|
2019-01-04 06:09:41 +00:00
|
|
|
#endif
|
2014-06-07 14:16:08 +01:00
|
|
|
|
2014-06-08 00:37:40 +01:00
|
|
|
mp_obj_t result;
|
2016-09-27 14:08:10 +01:00
|
|
|
if (vm_return_kind == MP_VM_RETURN_NORMAL) {
|
|
|
|
// return value is in *sp
|
|
|
|
result = *code_state->sp;
|
|
|
|
} else {
|
|
|
|
// must be an exception because normal functions can't yield
|
|
|
|
assert(vm_return_kind == MP_VM_RETURN_EXCEPTION);
|
2019-01-04 06:22:40 +00:00
|
|
|
// returned exception is in state[0]
|
2018-09-29 14:25:08 +01:00
|
|
|
result = code_state->state[0];
|
2014-06-07 14:16:08 +01:00
|
|
|
}
|
|
|
|
|
2017-11-26 12:37:19 +00:00
|
|
|
#if MICROPY_ENABLE_PYSTACK
|
|
|
|
mp_pystack_free(code_state);
|
|
|
|
#else
|
2014-06-07 14:16:08 +01:00
|
|
|
// free the state if it was allocated on the heap
|
2015-06-08 22:07:27 +01:00
|
|
|
if (state_size != 0) {
|
2016-08-27 14:21:00 +01:00
|
|
|
m_del_var(mp_code_state_t, byte, state_size, code_state);
|
2014-06-07 14:16:08 +01:00
|
|
|
}
|
2017-11-26 12:37:19 +00:00
|
|
|
#endif
|
2014-06-07 14:16:08 +01:00
|
|
|
|
2014-02-15 22:55:00 +00:00
|
|
|
if (vm_return_kind == MP_VM_RETURN_NORMAL) {
|
|
|
|
return result;
|
|
|
|
} else { // MP_VM_RETURN_EXCEPTION
|
2014-04-05 18:32:08 +01:00
|
|
|
nlr_raise(result);
|
2014-02-15 22:55:00 +00:00
|
|
|
}
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
|
2015-02-14 17:44:31 +00:00
|
|
|
#if MICROPY_PY_FUNCTION_ATTRS
|
2018-07-10 07:21:34 +01:00
|
|
|
void mp_obj_fun_bc_attr(mp_obj_t self_in, qstr attr, mp_obj_t *dest) {
|
2015-04-01 15:10:50 +01:00
|
|
|
if (dest[0] != MP_OBJ_NULL) {
|
|
|
|
// not load attribute
|
|
|
|
return;
|
|
|
|
}
|
2015-04-09 16:29:54 +01:00
|
|
|
if (attr == MP_QSTR___name__) {
|
2015-02-14 17:44:31 +00:00
|
|
|
dest[0] = MP_OBJ_NEW_QSTR(mp_obj_fun_get_name(self_in));
|
|
|
|
}
|
2020-09-25 16:17:11 +01:00
|
|
|
if (attr == MP_QSTR___globals__) {
|
|
|
|
mp_obj_fun_bc_t *self = MP_OBJ_TO_PTR(self_in);
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
dest[0] = MP_OBJ_FROM_PTR(self->context->module.globals);
|
2020-09-25 16:17:11 +01:00
|
|
|
}
|
2015-02-14 17:44:31 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2014-03-29 13:43:38 +00:00
|
|
|
const mp_obj_type_t mp_type_fun_bc = {
|
2014-02-15 16:10:44 +00:00
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF,
|
2014-02-15 11:34:50 +00:00
|
|
|
.name = MP_QSTR_function,
|
2014-04-30 23:32:58 +01:00
|
|
|
#if MICROPY_CPYTHON_COMPAT
|
|
|
|
.print = fun_bc_print,
|
|
|
|
#endif
|
2014-01-18 14:10:48 +00:00
|
|
|
.call = fun_bc_call,
|
2015-05-11 13:25:19 +01:00
|
|
|
.unary_op = mp_generic_unary_op,
|
2015-02-14 17:44:31 +00:00
|
|
|
#if MICROPY_PY_FUNCTION_ATTRS
|
2018-07-10 07:21:34 +01:00
|
|
|
.attr = mp_obj_fun_bc_attr,
|
2015-02-14 17:44:31 +00:00
|
|
|
#endif
|
2013-12-21 18:17:45 +00:00
|
|
|
};
|
|
|
|
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_obj_t mp_obj_new_fun_bc(const mp_obj_t *def_args, const byte *code, const mp_module_context_t *context, struct _mp_raw_code_t *const *child_table) {
|
2017-02-16 05:34:53 +00:00
|
|
|
size_t n_def_args = 0;
|
|
|
|
size_t n_extra_args = 0;
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_obj_tuple_t *def_pos_args = NULL;
|
|
|
|
mp_obj_t def_kw_args = MP_OBJ_NULL;
|
|
|
|
if (def_args != NULL && def_args[0] != MP_OBJ_NULL) {
|
|
|
|
assert(mp_obj_is_type(def_args[0], &mp_type_tuple));
|
|
|
|
def_pos_args = MP_OBJ_TO_PTR(def_args[0]);
|
|
|
|
n_def_args = def_pos_args->len;
|
|
|
|
n_extra_args = def_pos_args->len;
|
2014-02-16 00:01:29 +00:00
|
|
|
}
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
if (def_args != NULL && def_args[1] != MP_OBJ_NULL) {
|
|
|
|
assert(mp_obj_is_type(def_args[1], &mp_type_dict));
|
|
|
|
def_kw_args = def_args[1];
|
2014-06-07 22:01:00 +01:00
|
|
|
n_extra_args += 1;
|
|
|
|
}
|
2014-02-16 00:01:29 +00:00
|
|
|
mp_obj_fun_bc_t *o = m_new_obj_var(mp_obj_fun_bc_t, mp_obj_t, n_extra_args);
|
2014-03-29 13:43:38 +00:00
|
|
|
o->base.type = &mp_type_fun_bc;
|
2014-01-03 14:03:48 +00:00
|
|
|
o->bytecode = code;
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
o->context = context;
|
|
|
|
o->child_table = child_table;
|
|
|
|
if (def_pos_args != NULL) {
|
|
|
|
memcpy(o->extra_args, def_pos_args->items, n_def_args * sizeof(mp_obj_t));
|
2014-06-07 22:01:00 +01:00
|
|
|
}
|
|
|
|
if (def_kw_args != MP_OBJ_NULL) {
|
2014-06-08 00:37:40 +01:00
|
|
|
o->extra_args[n_def_args] = def_kw_args;
|
2014-04-27 15:50:52 +01:00
|
|
|
}
|
2015-11-27 17:01:44 +00:00
|
|
|
return MP_OBJ_FROM_PTR(o);
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
|
2014-08-24 16:28:17 +01:00
|
|
|
/******************************************************************************/
|
|
|
|
/* native functions */
|
|
|
|
|
|
|
|
#if MICROPY_EMIT_NATIVE
|
|
|
|
|
2016-01-03 09:59:18 +00:00
|
|
|
STATIC mp_obj_t fun_native_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2015-04-06 22:38:53 +01:00
|
|
|
MP_STACK_CHECK();
|
2022-01-11 16:25:06 +00:00
|
|
|
mp_obj_fun_bc_t *self = MP_OBJ_TO_PTR(self_in);
|
2015-04-06 22:38:53 +01:00
|
|
|
mp_call_fun_t fun = MICROPY_MAKE_POINTER_CALLABLE((void *)self->bytecode);
|
|
|
|
return fun(self_in, n_args, n_kw, args);
|
2014-08-24 16:28:17 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
STATIC const mp_obj_type_t mp_type_fun_native = {
|
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF,
|
2014-08-24 16:28:17 +01:00
|
|
|
.name = MP_QSTR_function,
|
|
|
|
.call = fun_native_call,
|
2015-05-11 13:25:19 +01:00
|
|
|
.unary_op = mp_generic_unary_op,
|
2014-08-24 16:28:17 +01:00
|
|
|
};
|
|
|
|
|
py: Rework bytecode and .mpy file format to be mostly static data.
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
|
|
|
mp_obj_t mp_obj_new_fun_native(const mp_obj_t *def_args, const void *fun_data, const mp_module_context_t *mc, struct _mp_raw_code_t *const *child_table) {
|
|
|
|
mp_obj_fun_bc_t *o = MP_OBJ_TO_PTR(mp_obj_new_fun_bc(def_args, (const byte *)fun_data, mc, child_table));
|
2014-08-24 16:28:17 +01:00
|
|
|
o->base.type = &mp_type_fun_native;
|
2022-01-11 16:25:06 +00:00
|
|
|
return MP_OBJ_FROM_PTR(o);
|
2014-08-24 16:28:17 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
#endif // MICROPY_EMIT_NATIVE
|
2014-08-15 16:45:41 +01:00
|
|
|
|
2013-12-21 18:17:45 +00:00
|
|
|
/******************************************************************************/
|
|
|
|
/* inline assembler functions */
|
|
|
|
|
2016-12-09 02:17:49 +00:00
|
|
|
#if MICROPY_EMIT_INLINE_ASM
|
2014-08-15 16:45:41 +01:00
|
|
|
|
2013-12-21 18:17:45 +00:00
|
|
|
typedef struct _mp_obj_fun_asm_t {
|
|
|
|
mp_obj_base_t base;
|
2017-02-16 05:34:53 +00:00
|
|
|
size_t n_args;
|
2019-02-20 02:08:42 +00:00
|
|
|
const void *fun_data; // GC must be able to trace this pointer
|
2016-01-15 15:20:43 +00:00
|
|
|
mp_uint_t type_sig;
|
2013-12-21 18:17:45 +00:00
|
|
|
} mp_obj_fun_asm_t;
|
|
|
|
|
2015-01-12 22:45:35 +00:00
|
|
|
typedef mp_uint_t (*inline_asm_fun_0_t)(void);
|
2014-07-03 13:25:24 +01:00
|
|
|
typedef mp_uint_t (*inline_asm_fun_1_t)(mp_uint_t);
|
|
|
|
typedef mp_uint_t (*inline_asm_fun_2_t)(mp_uint_t, mp_uint_t);
|
|
|
|
typedef mp_uint_t (*inline_asm_fun_3_t)(mp_uint_t, mp_uint_t, mp_uint_t);
|
2016-03-16 08:22:26 +00:00
|
|
|
typedef mp_uint_t (*inline_asm_fun_4_t)(mp_uint_t, mp_uint_t, mp_uint_t, mp_uint_t);
|
2013-12-21 18:17:45 +00:00
|
|
|
|
2017-06-30 08:22:17 +01:00
|
|
|
// convert a MicroPython object to a sensible value for inline asm
|
2014-07-03 13:25:24 +01:00
|
|
|
STATIC mp_uint_t convert_obj_for_inline_asm(mp_obj_t obj) {
|
2013-12-21 18:17:45 +00:00
|
|
|
// TODO for byte_array, pass pointer to the array
|
2019-01-30 07:49:52 +00:00
|
|
|
if (mp_obj_is_small_int(obj)) {
|
2013-12-21 18:17:45 +00:00
|
|
|
return MP_OBJ_SMALL_INT_VALUE(obj);
|
|
|
|
} else if (obj == mp_const_none) {
|
|
|
|
return 0;
|
|
|
|
} else if (obj == mp_const_false) {
|
|
|
|
return 0;
|
|
|
|
} else if (obj == mp_const_true) {
|
|
|
|
return 1;
|
2019-01-30 07:49:52 +00:00
|
|
|
} else if (mp_obj_is_type(obj, &mp_type_int)) {
|
2015-02-13 10:43:05 +00:00
|
|
|
return mp_obj_int_get_truncated(obj);
|
2019-01-30 07:49:52 +00:00
|
|
|
} else if (mp_obj_is_str(obj)) {
|
2013-12-21 18:17:45 +00:00
|
|
|
// pointer to the string (it's probably constant though!)
|
2017-03-25 08:48:18 +00:00
|
|
|
size_t l;
|
2014-07-03 13:25:24 +01:00
|
|
|
return (mp_uint_t)mp_obj_str_get_data(obj, &l);
|
2014-04-13 00:30:32 +01:00
|
|
|
} else {
|
2020-01-09 00:01:14 +00:00
|
|
|
const mp_obj_type_t *type = mp_obj_get_type(obj);
|
2020-02-19 13:17:13 +00:00
|
|
|
#if MICROPY_PY_BUILTINS_FLOAT
|
2019-05-05 04:29:43 +01:00
|
|
|
if (type == &mp_type_float) {
|
2014-04-13 00:30:32 +01:00
|
|
|
// convert float to int (could also pass in float registers)
|
2014-07-03 13:25:24 +01:00
|
|
|
return (mp_int_t)mp_obj_float_get(obj);
|
2020-02-19 13:17:13 +00:00
|
|
|
}
|
|
|
|
#endif
|
2019-05-05 04:29:43 +01:00
|
|
|
if (type == &mp_type_tuple || type == &mp_type_list) {
|
2014-04-13 00:30:32 +01:00
|
|
|
// pointer to start of tuple (could pass length, but then could use len(x) for that)
|
2017-03-25 08:35:08 +00:00
|
|
|
size_t len;
|
2014-04-13 00:30:32 +01:00
|
|
|
mp_obj_t *items;
|
2017-03-04 11:29:20 +00:00
|
|
|
mp_obj_get_array(obj, &len, &items);
|
2014-07-03 13:25:24 +01:00
|
|
|
return (mp_uint_t)items;
|
2014-04-13 00:30:32 +01:00
|
|
|
} else {
|
2014-04-18 22:29:21 +01:00
|
|
|
mp_buffer_info_t bufinfo;
|
2019-08-05 07:32:10 +01:00
|
|
|
if (mp_get_buffer(obj, &bufinfo, MP_BUFFER_READ)) {
|
2014-04-13 12:08:52 +01:00
|
|
|
// supports the buffer protocol, return a pointer to the data
|
2014-07-03 13:25:24 +01:00
|
|
|
return (mp_uint_t)bufinfo.buf;
|
2014-04-13 12:08:52 +01:00
|
|
|
} else {
|
|
|
|
// just pass along a pointer to the object
|
2014-07-03 13:25:24 +01:00
|
|
|
return (mp_uint_t)obj;
|
2014-04-13 12:08:52 +01:00
|
|
|
}
|
2014-04-13 00:30:32 +01:00
|
|
|
}
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-01-03 09:59:18 +00:00
|
|
|
STATIC mp_obj_t fun_asm_call(mp_obj_t self_in, size_t n_args, size_t n_kw, const mp_obj_t *args) {
|
2022-01-11 16:25:06 +00:00
|
|
|
mp_obj_fun_asm_t *self = MP_OBJ_TO_PTR(self_in);
|
2013-12-21 18:17:45 +00:00
|
|
|
|
2014-05-10 13:40:46 +01:00
|
|
|
mp_arg_check_num(n_args, n_kw, self->n_args, self->n_args, false);
|
2013-12-21 18:17:45 +00:00
|
|
|
|
2019-02-20 02:08:42 +00:00
|
|
|
const void *fun = MICROPY_MAKE_POINTER_CALLABLE(self->fun_data);
|
2014-08-24 16:28:17 +01:00
|
|
|
|
2014-07-03 13:25:24 +01:00
|
|
|
mp_uint_t ret;
|
2013-12-21 18:17:45 +00:00
|
|
|
if (n_args == 0) {
|
2014-08-24 16:28:17 +01:00
|
|
|
ret = ((inline_asm_fun_0_t)fun)();
|
2013-12-21 18:17:45 +00:00
|
|
|
} else if (n_args == 1) {
|
2014-08-24 16:28:17 +01:00
|
|
|
ret = ((inline_asm_fun_1_t)fun)(convert_obj_for_inline_asm(args[0]));
|
2013-12-21 18:17:45 +00:00
|
|
|
} else if (n_args == 2) {
|
2014-08-24 16:28:17 +01:00
|
|
|
ret = ((inline_asm_fun_2_t)fun)(convert_obj_for_inline_asm(args[0]), convert_obj_for_inline_asm(args[1]));
|
2013-12-21 18:17:45 +00:00
|
|
|
} else if (n_args == 3) {
|
2014-08-24 16:28:17 +01:00
|
|
|
ret = ((inline_asm_fun_3_t)fun)(convert_obj_for_inline_asm(args[0]), convert_obj_for_inline_asm(args[1]), convert_obj_for_inline_asm(args[2]));
|
2013-12-21 18:17:45 +00:00
|
|
|
} else {
|
2016-03-16 08:22:26 +00:00
|
|
|
// compiler allows at most 4 arguments
|
|
|
|
assert(n_args == 4);
|
|
|
|
ret = ((inline_asm_fun_4_t)fun)(
|
|
|
|
convert_obj_for_inline_asm(args[0]),
|
|
|
|
convert_obj_for_inline_asm(args[1]),
|
|
|
|
convert_obj_for_inline_asm(args[2]),
|
|
|
|
convert_obj_for_inline_asm(args[3])
|
|
|
|
);
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
|
2019-03-08 23:22:09 +00:00
|
|
|
return mp_native_to_obj(ret, self->type_sig);
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
|
|
|
|
2014-03-29 13:43:38 +00:00
|
|
|
STATIC const mp_obj_type_t mp_type_fun_asm = {
|
2014-02-15 16:10:44 +00:00
|
|
|
{ &mp_type_type },
|
2020-06-28 00:39:20 +01:00
|
|
|
.flags = MP_TYPE_FLAG_BINDS_SELF,
|
2014-02-15 11:34:50 +00:00
|
|
|
.name = MP_QSTR_function,
|
2014-01-18 14:10:48 +00:00
|
|
|
.call = fun_asm_call,
|
2015-05-11 13:25:19 +01:00
|
|
|
.unary_op = mp_generic_unary_op,
|
2013-12-21 18:17:45 +00:00
|
|
|
};
|
|
|
|
|
2019-02-20 02:08:42 +00:00
|
|
|
mp_obj_t mp_obj_new_fun_asm(size_t n_args, const void *fun_data, mp_uint_t type_sig) {
|
2013-12-21 18:17:45 +00:00
|
|
|
mp_obj_fun_asm_t *o = m_new_obj(mp_obj_fun_asm_t);
|
2014-03-29 13:43:38 +00:00
|
|
|
o->base.type = &mp_type_fun_asm;
|
2013-12-21 18:17:45 +00:00
|
|
|
o->n_args = n_args;
|
2014-08-24 16:28:17 +01:00
|
|
|
o->fun_data = fun_data;
|
2016-01-15 15:20:43 +00:00
|
|
|
o->type_sig = type_sig;
|
2022-01-11 16:25:06 +00:00
|
|
|
return MP_OBJ_FROM_PTR(o);
|
2013-12-21 18:17:45 +00:00
|
|
|
}
|
2014-08-15 16:45:41 +01:00
|
|
|
|
2016-12-09 02:17:49 +00:00
|
|
|
#endif // MICROPY_EMIT_INLINE_ASM
|