2014-05-03 18:27:38 -04:00
|
|
|
/*
|
2017-06-30 03:22:17 -04:00
|
|
|
* This file is part of the MicroPython project, http://micropython.org/
|
2014-05-03 18:27:38 -04:00
|
|
|
*
|
|
|
|
* The MIT License (MIT)
|
|
|
|
*
|
2019-08-22 00:55:16 -04:00
|
|
|
* Copyright (c) 2013-2019 Damien P. George
|
2014-05-03 18:27:38 -04:00
|
|
|
*
|
|
|
|
* 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.
|
|
|
|
*/
|
|
|
|
|
2014-03-16 03:14:26 -04:00
|
|
|
#include <stdbool.h>
|
2013-10-04 14:53:11 -04:00
|
|
|
#include <stdint.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
#include <string.h>
|
2022-03-15 18:37:58 -04:00
|
|
|
#include <unistd.h>
|
2013-10-04 14:53:11 -04:00
|
|
|
#include <assert.h>
|
|
|
|
|
2015-01-01 18:30:53 -05:00
|
|
|
#include "py/mpstate.h"
|
2022-03-31 08:59:10 -04:00
|
|
|
#include "py/smallint.h"
|
2015-01-01 15:27:54 -05:00
|
|
|
#include "py/emit.h"
|
|
|
|
#include "py/bc0.h"
|
2013-10-04 14:53:11 -04:00
|
|
|
|
2015-12-18 07:35:44 -05:00
|
|
|
#if MICROPY_ENABLE_COMPILER
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
#define DUMMY_DATA_SIZE (MP_ENCODE_UINT_MAX_BYTES)
|
2014-05-10 13:07:08 -04:00
|
|
|
|
2013-10-05 07:19:06 -04:00
|
|
|
struct _emit_t {
|
2015-11-27 12:01:44 -05:00
|
|
|
// Accessed as mp_obj_t, so must be aligned as such, and we rely on the
|
|
|
|
// memory allocator returning a suitably aligned pointer.
|
|
|
|
// Should work for cases when mp_obj_t is 64-bit on a 32-bit machine.
|
|
|
|
byte dummy_data[DUMMY_DATA_SIZE];
|
|
|
|
|
2014-05-10 13:16:21 -04:00
|
|
|
pass_kind_t pass : 8;
|
|
|
|
|
2022-06-17 09:14:32 -04:00
|
|
|
// Set to true if the code generator should suppress emitted code due to it
|
|
|
|
// being dead code. This can happen when opcodes immediately follow an
|
|
|
|
// unconditional flow control (eg jump or raise).
|
|
|
|
bool suppress;
|
|
|
|
|
2013-10-04 14:53:11 -04:00
|
|
|
int stack_size;
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
mp_emit_common_t *emit_common;
|
2013-10-04 14:53:11 -04:00
|
|
|
scope_t *scope;
|
|
|
|
|
2014-09-08 18:05:16 -04:00
|
|
|
mp_uint_t last_source_line_offset;
|
|
|
|
mp_uint_t last_source_line;
|
2014-01-18 18:24:36 -05:00
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
size_t max_num_labels;
|
|
|
|
size_t *label_offsets;
|
2013-10-04 14:53:11 -04:00
|
|
|
|
2015-11-27 12:01:44 -05:00
|
|
|
size_t code_info_offset;
|
|
|
|
size_t code_info_size;
|
|
|
|
size_t bytecode_offset;
|
|
|
|
size_t bytecode_size;
|
2014-01-18 18:24:36 -05:00
|
|
|
byte *code_base; // stores both byte code and code info
|
2022-03-17 22:18:59 -04:00
|
|
|
bool overflow;
|
2015-11-02 12:27:18 -05:00
|
|
|
|
2019-09-25 01:45:47 -04:00
|
|
|
size_t n_info;
|
|
|
|
size_t n_cell;
|
2013-10-04 14:53:11 -04:00
|
|
|
};
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
emit_t *emit_bc_new(mp_emit_common_t *emit_common) {
|
2014-01-18 18:24:36 -05:00
|
|
|
emit_t *emit = m_new0(emit_t, 1);
|
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 07:22:47 -04:00
|
|
|
emit->emit_common = emit_common;
|
2015-03-26 11:49:53 -04:00
|
|
|
return emit;
|
|
|
|
}
|
|
|
|
|
2015-04-09 11:29:54 -04:00
|
|
|
void emit_bc_set_max_num_labels(emit_t *emit, mp_uint_t max_num_labels) {
|
2013-10-05 13:08:26 -04:00
|
|
|
emit->max_num_labels = max_num_labels;
|
2022-03-15 18:37:58 -04:00
|
|
|
emit->label_offsets = m_new(size_t, emit->max_num_labels);
|
2013-10-05 13:08:26 -04:00
|
|
|
}
|
2013-10-05 09:17:09 -04:00
|
|
|
|
2014-01-24 17:42:28 -05:00
|
|
|
void emit_bc_free(emit_t *emit) {
|
2022-03-15 18:37:58 -04:00
|
|
|
m_del(size_t, emit->label_offsets, emit->max_num_labels);
|
2014-01-24 09:20:11 -05:00
|
|
|
m_del_obj(emit_t, emit);
|
|
|
|
}
|
|
|
|
|
2014-01-18 18:24:36 -05:00
|
|
|
// all functions must go through this one to emit 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 07:22:47 -04:00
|
|
|
STATIC uint8_t *emit_get_cur_to_write_code_info(void *emit_in, size_t num_bytes_to_write) {
|
|
|
|
emit_t *emit = emit_in;
|
2014-05-07 12:24:22 -04:00
|
|
|
if (emit->pass < MP_PASS_EMIT) {
|
2014-01-18 18:24:36 -05:00
|
|
|
emit->code_info_offset += num_bytes_to_write;
|
|
|
|
return emit->dummy_data;
|
|
|
|
} else {
|
|
|
|
assert(emit->code_info_offset + num_bytes_to_write <= emit->code_info_size);
|
|
|
|
byte *c = emit->code_base + emit->code_info_offset;
|
|
|
|
emit->code_info_offset += num_bytes_to_write;
|
|
|
|
return c;
|
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2020-02-26 23:36:53 -05:00
|
|
|
STATIC void emit_write_code_info_byte(emit_t *emit, byte val) {
|
2015-03-18 13:47:47 -04:00
|
|
|
*emit_get_cur_to_write_code_info(emit, 1) = val;
|
|
|
|
}
|
|
|
|
|
2015-04-09 11:29:54 -04:00
|
|
|
STATIC void emit_write_code_info_qstr(emit_t *emit, qstr qst) {
|
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 07:22:47 -04:00
|
|
|
mp_encode_uint(emit, emit_get_cur_to_write_code_info, mp_emit_common_use_qstr(emit->emit_common, qst));
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2014-04-13 09:51:56 -04:00
|
|
|
#if MICROPY_ENABLE_SOURCE_LINE
|
2015-04-09 11:29:54 -04:00
|
|
|
STATIC void emit_write_code_info_bytes_lines(emit_t *emit, mp_uint_t bytes_to_skip, mp_uint_t lines_to_skip) {
|
2014-04-13 09:51:56 -04:00
|
|
|
assert(bytes_to_skip > 0 || lines_to_skip > 0);
|
|
|
|
while (bytes_to_skip > 0 || lines_to_skip > 0) {
|
2014-07-31 12:12:01 -04:00
|
|
|
mp_uint_t b, l;
|
2017-02-09 19:58:10 -05:00
|
|
|
if (lines_to_skip <= 6 || bytes_to_skip > 0xf) {
|
2014-07-31 12:12:01 -04:00
|
|
|
// use 0b0LLBBBBB encoding
|
|
|
|
b = MIN(bytes_to_skip, 0x1f);
|
2017-02-09 19:58:10 -05:00
|
|
|
if (b < bytes_to_skip) {
|
|
|
|
// we can't skip any lines until we skip all the bytes
|
|
|
|
l = 0;
|
|
|
|
} else {
|
|
|
|
l = MIN(lines_to_skip, 0x3);
|
|
|
|
}
|
2014-07-31 12:12:01 -04:00
|
|
|
*emit_get_cur_to_write_code_info(emit, 1) = b | (l << 5);
|
|
|
|
} else {
|
|
|
|
// use 0b1LLLBBBB 0bLLLLLLLL encoding (l's LSB in second byte)
|
|
|
|
b = MIN(bytes_to_skip, 0xf);
|
|
|
|
l = MIN(lines_to_skip, 0x7ff);
|
|
|
|
byte *ci = emit_get_cur_to_write_code_info(emit, 2);
|
|
|
|
ci[0] = 0x80 | b | ((l >> 4) & 0x70);
|
|
|
|
ci[1] = l;
|
|
|
|
}
|
2014-04-13 09:51:56 -04:00
|
|
|
bytes_to_skip -= b;
|
|
|
|
lines_to_skip -= l;
|
2014-01-25 06:43:20 -05:00
|
|
|
}
|
2014-01-18 18:24:36 -05:00
|
|
|
}
|
2014-04-13 09:51:56 -04:00
|
|
|
#endif
|
2014-01-18 18:24:36 -05:00
|
|
|
|
|
|
|
// all functions must go through this one to emit byte 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 07:22:47 -04:00
|
|
|
STATIC uint8_t *emit_get_cur_to_write_bytecode(void *emit_in, size_t num_bytes_to_write) {
|
|
|
|
emit_t *emit = emit_in;
|
2022-06-17 09:14:32 -04:00
|
|
|
if (emit->suppress) {
|
|
|
|
return emit->dummy_data;
|
|
|
|
}
|
2014-05-07 12:24:22 -04:00
|
|
|
if (emit->pass < MP_PASS_EMIT) {
|
2014-05-10 05:36:38 -04:00
|
|
|
emit->bytecode_offset += num_bytes_to_write;
|
2013-10-04 14:53:11 -04:00
|
|
|
return emit->dummy_data;
|
|
|
|
} else {
|
2014-05-10 05:36:38 -04:00
|
|
|
assert(emit->bytecode_offset + num_bytes_to_write <= emit->bytecode_size);
|
|
|
|
byte *c = emit->code_base + emit->code_info_size + emit->bytecode_offset;
|
|
|
|
emit->bytecode_offset += num_bytes_to_write;
|
2013-10-04 14:53:11 -04:00
|
|
|
return c;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_write_bytecode_raw_byte(emit_t *emit, byte b1) {
|
2015-04-09 11:29:54 -04:00
|
|
|
byte *c = emit_get_cur_to_write_bytecode(emit, 1);
|
2013-10-04 14:53:11 -04:00
|
|
|
c[0] = b1;
|
|
|
|
}
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_write_bytecode_byte(emit_t *emit, int stack_adj, byte b1) {
|
|
|
|
mp_emit_bc_adjust_stack_size(emit, stack_adj);
|
|
|
|
byte *c = emit_get_cur_to_write_bytecode(emit, 1);
|
|
|
|
c[0] = b1;
|
|
|
|
}
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
// Similar to mp_encode_uint(), just some extra handling to encode sign
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_write_bytecode_byte_int(emit_t *emit, int stack_adj, byte b1, mp_int_t num) {
|
|
|
|
emit_write_bytecode_byte(emit, stack_adj, b1);
|
2014-02-19 08:47:59 -05:00
|
|
|
|
|
|
|
// We store each 7 bits in a separate byte, and that's how many bytes needed
|
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 07:22:47 -04:00
|
|
|
byte buf[MP_ENCODE_UINT_MAX_BYTES];
|
2014-02-19 08:47:59 -05:00
|
|
|
byte *p = buf + sizeof(buf);
|
|
|
|
// We encode in little-ending order, but store in big-endian, to help decoding
|
|
|
|
do {
|
|
|
|
*--p = num & 0x7f;
|
|
|
|
num >>= 7;
|
|
|
|
} while (num != 0 && num != -1);
|
|
|
|
// Make sure that highest bit we stored (mask 0x40) matches sign
|
|
|
|
// of the number. If not, store extra byte just to encode sign
|
|
|
|
if (num == -1 && (*p & 0x40) == 0) {
|
|
|
|
*--p = 0x7f;
|
|
|
|
} else if (num == 0 && (*p & 0x40) != 0) {
|
|
|
|
*--p = 0;
|
|
|
|
}
|
|
|
|
|
2015-04-09 11:29:54 -04:00
|
|
|
byte *c = emit_get_cur_to_write_bytecode(emit, buf + sizeof(buf) - p);
|
2014-02-19 08:47:59 -05:00
|
|
|
while (p != buf + sizeof(buf) - 1) {
|
|
|
|
*c++ = *p++ | 0x80;
|
|
|
|
}
|
|
|
|
*c = *p;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_write_bytecode_byte_uint(emit_t *emit, int stack_adj, byte b, mp_uint_t val) {
|
|
|
|
emit_write_bytecode_byte(emit, stack_adj, b);
|
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 07:22:47 -04:00
|
|
|
mp_encode_uint(emit, emit_get_cur_to_write_bytecode, val);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
STATIC void emit_write_bytecode_byte_const(emit_t *emit, int stack_adj, byte b, mp_uint_t n) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, stack_adj, b, n);
|
2015-11-02 12:27:18 -05:00
|
|
|
}
|
2014-04-13 06:04:33 -04:00
|
|
|
|
2020-02-26 23:36:53 -05:00
|
|
|
STATIC void emit_write_bytecode_byte_qstr(emit_t *emit, int stack_adj, byte b, qstr qst) {
|
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 07:22:47 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, stack_adj, b, mp_emit_common_use_qstr(emit->emit_common, qst));
|
2015-11-02 12:27:18 -05:00
|
|
|
}
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_write_bytecode_byte_obj(emit_t *emit, int stack_adj, byte b, mp_obj_t obj) {
|
2022-05-07 01:51:41 -04:00
|
|
|
emit_write_bytecode_byte_const(emit, stack_adj, b, mp_emit_common_use_const_obj(emit->emit_common, obj));
|
2015-11-02 12:27:18 -05:00
|
|
|
}
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
STATIC void emit_write_bytecode_byte_child(emit_t *emit, int stack_adj, byte b, mp_raw_code_t *rc) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_const(emit, stack_adj, b,
|
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 07:22:47 -04:00
|
|
|
mp_emit_common_alloc_const_child(emit->emit_common, rc));
|
2019-08-14 10:09:36 -04:00
|
|
|
#if MICROPY_PY_SYS_SETTRACE
|
|
|
|
rc->line_of_definition = emit->last_source_line;
|
|
|
|
#endif
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
// Emit a jump opcode to a destination label.
|
|
|
|
// The offset to the label is relative to the ip following this instruction.
|
|
|
|
// The offset is encoded as either 1 or 2 bytes, depending on how big it is.
|
|
|
|
// The encoding of this jump opcode can change size from one pass to the next,
|
|
|
|
// but it must only ever decrease in size on successive passes.
|
|
|
|
STATIC void emit_write_bytecode_byte_label(emit_t *emit, int stack_adj, byte b1, mp_uint_t label) {
|
2019-08-22 00:55:16 -04:00
|
|
|
mp_emit_bc_adjust_stack_size(emit, stack_adj);
|
2022-03-15 18:37:58 -04:00
|
|
|
|
2022-06-17 09:14:32 -04:00
|
|
|
if (emit->suppress) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
// Determine if the jump offset is signed or unsigned, based on the opcode.
|
2022-03-21 01:36:13 -04:00
|
|
|
const bool is_signed = b1 <= MP_BC_POP_JUMP_IF_FALSE;
|
2022-03-15 18:37:58 -04:00
|
|
|
|
|
|
|
// Default to a 2-byte encoding (the largest) with an unknown jump offset.
|
|
|
|
unsigned int jump_encoding_size = 1;
|
|
|
|
ssize_t bytecode_offset = 0;
|
|
|
|
|
|
|
|
// Compute the jump size and offset only when code size is known.
|
|
|
|
if (emit->pass >= MP_PASS_CODE_SIZE) {
|
|
|
|
// The -2 accounts for this jump opcode taking 2 bytes (at least).
|
|
|
|
bytecode_offset = emit->label_offsets[label] - emit->bytecode_offset - 2;
|
|
|
|
|
|
|
|
// Check if the bytecode_offset is small enough to use a 1-byte encoding.
|
|
|
|
if ((is_signed && -64 <= bytecode_offset && bytecode_offset <= 63)
|
|
|
|
|| (!is_signed && (size_t)bytecode_offset <= 127)) {
|
|
|
|
// Use a 1-byte jump offset.
|
|
|
|
jump_encoding_size = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Adjust the offset depending on the size of the encoding of the offset.
|
|
|
|
bytecode_offset -= jump_encoding_size;
|
|
|
|
|
|
|
|
assert(is_signed || bytecode_offset >= 0);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
2013-11-05 17:06:08 -05:00
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
// Emit the opcode.
|
|
|
|
byte *c = emit_get_cur_to_write_bytecode(emit, 2 + jump_encoding_size);
|
|
|
|
c[0] = b1;
|
|
|
|
if (jump_encoding_size == 0) {
|
|
|
|
if (is_signed) {
|
|
|
|
bytecode_offset += 0x40;
|
|
|
|
}
|
|
|
|
assert(0 <= bytecode_offset && bytecode_offset <= 0x7f);
|
|
|
|
c[1] = bytecode_offset;
|
2013-11-05 17:06:08 -05:00
|
|
|
} else {
|
2022-03-15 18:37:58 -04:00
|
|
|
if (is_signed) {
|
|
|
|
bytecode_offset += 0x4000;
|
|
|
|
}
|
2022-03-17 22:18:59 -04:00
|
|
|
if (emit->pass == MP_PASS_EMIT && !(0 <= bytecode_offset && bytecode_offset <= 0x7fff)) {
|
|
|
|
emit->overflow = true;
|
|
|
|
}
|
2022-03-15 18:37:58 -04:00
|
|
|
c[1] = 0x80 | (bytecode_offset & 0x7f);
|
|
|
|
c[2] = bytecode_offset >> 7;
|
2013-11-05 17:06:08 -05:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_start_pass(emit_t *emit, pass_kind_t pass, scope_t *scope) {
|
2013-12-30 17:32:17 -05:00
|
|
|
emit->pass = pass;
|
|
|
|
emit->stack_size = 0;
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = false;
|
2013-12-30 17:32:17 -05:00
|
|
|
emit->scope = scope;
|
2014-01-18 18:24:36 -05:00
|
|
|
emit->last_source_line_offset = 0;
|
|
|
|
emit->last_source_line = 1;
|
2014-05-10 05:36:38 -04:00
|
|
|
emit->bytecode_offset = 0;
|
2014-01-18 18:24:36 -05:00
|
|
|
emit->code_info_offset = 0;
|
2022-03-17 22:18:59 -04:00
|
|
|
emit->overflow = false;
|
2014-01-18 18:24:36 -05:00
|
|
|
|
2019-09-16 08:12:59 -04:00
|
|
|
// Write local state size, exception stack size, scope flags and number of arguments
|
2015-03-18 13:47:47 -04:00
|
|
|
{
|
|
|
|
mp_uint_t n_state = scope->num_locals + scope->stack_size;
|
|
|
|
if (n_state == 0) {
|
|
|
|
// Need at least 1 entry in the state, in the case an exception is
|
|
|
|
// propagated through this function, the exception is returned in
|
|
|
|
// the highest slot in the state (fastn[0], see vm.c).
|
|
|
|
n_state = 1;
|
|
|
|
}
|
2019-01-04 01:09:41 -05:00
|
|
|
#if MICROPY_DEBUG_VM_STACK_OVERFLOW
|
|
|
|
// An extra slot in the stack is needed to detect VM stack overflow
|
|
|
|
n_state += 1;
|
|
|
|
#endif
|
2014-01-18 18:24:36 -05:00
|
|
|
|
2019-09-16 08:12:59 -04:00
|
|
|
size_t n_exc_stack = scope->exc_stack_size;
|
|
|
|
MP_BC_PRELUDE_SIG_ENCODE(n_state, n_exc_stack, scope, emit_write_code_info_byte, emit);
|
|
|
|
}
|
2015-10-22 18:45:37 -04:00
|
|
|
|
2019-09-25 01:45:47 -04:00
|
|
|
// Write number of cells and size of the source 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 07:22:47 -04:00
|
|
|
if (emit->pass >= MP_PASS_CODE_SIZE) {
|
|
|
|
size_t n_info = emit->n_info;
|
|
|
|
size_t n_cell = emit->n_cell;
|
|
|
|
MP_BC_PRELUDE_SIZE_ENCODE(n_info, n_cell, emit_write_code_info_byte, emit);
|
2014-03-27 06:55:21 -04:00
|
|
|
}
|
|
|
|
|
2019-09-25 01:45:47 -04:00
|
|
|
emit->n_info = emit->code_info_offset;
|
|
|
|
|
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 07:22:47 -04:00
|
|
|
// Write the name of this function.
|
2015-03-18 13:47:47 -04:00
|
|
|
emit_write_code_info_qstr(emit, scope->simple_name);
|
2015-10-22 20:23:11 -04:00
|
|
|
|
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 07:22:47 -04:00
|
|
|
// Write argument names, needed to resolve positional args passed as keywords.
|
|
|
|
{
|
2015-10-22 20:23:11 -04:00
|
|
|
// For a given argument position (indexed by i) we need to find the
|
|
|
|
// corresponding id_info which is a parameter, as it has the correct
|
|
|
|
// qstr name to use as the argument name. Note that it's not a simple
|
|
|
|
// 1-1 mapping (ie i!=j in general) because of possible closed-over
|
|
|
|
// variables. In the case that the argument i has no corresponding
|
|
|
|
// parameter we use "*" as its name (since no argument can ever be named
|
|
|
|
// "*"). We could use a blank qstr but "*" is better for debugging.
|
|
|
|
// Note: there is some wasted RAM here for the case of storing a qstr
|
|
|
|
// for each closed-over variable, and maybe there is a better way to do
|
|
|
|
// it, but that would require changes to mp_setup_code_state.
|
|
|
|
for (int i = 0; i < scope->num_pos_args + scope->num_kwonly_args; i++) {
|
|
|
|
qstr qst = MP_QSTR__star_;
|
|
|
|
for (int j = 0; j < scope->id_info_len; ++j) {
|
|
|
|
id_info_t *id = &scope->id_info[j];
|
|
|
|
if ((id->flags & ID_FLAG_IS_PARAM) && id->local_num == i) {
|
|
|
|
qst = id->qst;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
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 07:22:47 -04:00
|
|
|
emit_write_code_info_qstr(emit, qst);
|
2015-10-22 20:23:11 -04:00
|
|
|
}
|
|
|
|
}
|
2013-12-30 17:32:17 -05:00
|
|
|
}
|
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
bool mp_emit_bc_end_pass(emit_t *emit) {
|
2015-03-26 11:49:53 -04:00
|
|
|
if (emit->pass == MP_PASS_SCOPE) {
|
2022-03-15 18:37:58 -04:00
|
|
|
return true;
|
2015-03-26 11:49:53 -04:00
|
|
|
}
|
|
|
|
|
2013-12-30 17:32:17 -05:00
|
|
|
// check stack is back to zero size
|
2016-09-27 01:46:50 -04:00
|
|
|
assert(emit->stack_size == 0);
|
2013-12-30 17:32:17 -05:00
|
|
|
|
2019-09-25 01:45:47 -04:00
|
|
|
// Calculate size of source code info section
|
|
|
|
emit->n_info = emit->code_info_offset - emit->n_info;
|
|
|
|
|
|
|
|
// Emit closure section of prelude
|
|
|
|
emit->n_cell = 0;
|
|
|
|
for (size_t i = 0; i < emit->scope->id_info_len; ++i) {
|
|
|
|
id_info_t *id = &emit->scope->id_info[i];
|
|
|
|
if (id->kind == ID_INFO_KIND_CELL) {
|
|
|
|
assert(id->local_num <= 255);
|
|
|
|
emit_write_code_info_byte(emit, id->local_num); // write the local which should be converted to a cell
|
|
|
|
++emit->n_cell;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-05-07 12:24:22 -04:00
|
|
|
if (emit->pass == MP_PASS_CODE_SIZE) {
|
2014-09-04 09:44:01 -04:00
|
|
|
// calculate size of total code-info + bytecode, in bytes
|
2014-01-18 18:24:36 -05:00
|
|
|
emit->code_info_size = emit->code_info_offset;
|
2014-05-10 05:36:38 -04:00
|
|
|
emit->bytecode_size = emit->bytecode_offset;
|
|
|
|
emit->code_base = m_new0(byte, emit->code_info_size + emit->bytecode_size);
|
2013-12-30 17:32:17 -05:00
|
|
|
|
2014-05-07 12:24:22 -04:00
|
|
|
} else if (emit->pass == MP_PASS_EMIT) {
|
2022-03-15 18:37:58 -04:00
|
|
|
// Code info and/or bytecode can shrink during this pass.
|
|
|
|
assert(emit->code_info_offset <= emit->code_info_size);
|
|
|
|
assert(emit->bytecode_offset <= emit->bytecode_size);
|
|
|
|
|
|
|
|
if (emit->code_info_offset != emit->code_info_size
|
|
|
|
|| emit->bytecode_offset != emit->bytecode_size) {
|
|
|
|
// Code info and/or bytecode changed size in this pass, so request the
|
|
|
|
// compiler to do another pass with these updated sizes.
|
|
|
|
emit->code_info_size = emit->code_info_offset;
|
|
|
|
emit->bytecode_size = emit->bytecode_offset;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2022-03-17 22:18:59 -04:00
|
|
|
if (emit->overflow) {
|
|
|
|
mp_raise_msg(&mp_type_RuntimeError, MP_ERROR_TEXT("bytecode overflow"));
|
|
|
|
}
|
|
|
|
|
2022-03-15 18:37:58 -04:00
|
|
|
// Bytecode is finalised, assign it to the raw code object.
|
2014-05-10 05:36:38 -04:00
|
|
|
mp_emit_glue_assign_bytecode(emit->scope->raw_code, emit->code_base,
|
2018-02-14 02:41:17 -05:00
|
|
|
#if MICROPY_PERSISTENT_CODE_SAVE || MICROPY_DEBUG_PRINTERS
|
2015-10-22 20:23:11 -04:00
|
|
|
emit->code_info_size + emit->bytecode_size,
|
2018-02-14 02:41:17 -05:00
|
|
|
#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 07:22:47 -04:00
|
|
|
emit->emit_common->children,
|
2015-11-02 16:55:42 -05:00
|
|
|
#if MICROPY_PERSISTENT_CODE_SAVE
|
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 07:22:47 -04:00
|
|
|
emit->emit_common->ct_cur_child,
|
2015-11-02 16:55:42 -05:00
|
|
|
#endif
|
|
|
|
emit->scope->scope_flags);
|
2013-12-30 17:32:17 -05:00
|
|
|
}
|
2022-03-15 18:37:58 -04:00
|
|
|
|
|
|
|
return true;
|
2013-12-30 17:32:17 -05:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_adjust_stack_size(emit_t *emit, mp_int_t delta) {
|
2017-05-25 06:42:30 -04:00
|
|
|
if (emit->pass == MP_PASS_SCOPE) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
assert((mp_int_t)emit->stack_size + delta >= 0);
|
2014-04-10 13:28:54 -04:00
|
|
|
emit->stack_size += delta;
|
2017-05-25 06:42:30 -04:00
|
|
|
if (emit->stack_size > emit->scope->stack_size) {
|
|
|
|
emit->scope->stack_size = emit->stack_size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_set_source_line(emit_t *emit, mp_uint_t source_line) {
|
2020-02-26 23:36:53 -05:00
|
|
|
#if MICROPY_ENABLE_SOURCE_LINE
|
2015-01-01 18:30:53 -05:00
|
|
|
if (MP_STATE_VM(mp_optimise_value) >= 3) {
|
2014-06-02 12:39:15 -04:00
|
|
|
// If we compile with -O3, don't store line numbers.
|
|
|
|
return;
|
|
|
|
}
|
2014-01-18 18:24:36 -05:00
|
|
|
if (source_line > emit->last_source_line) {
|
2014-09-08 18:05:16 -04:00
|
|
|
mp_uint_t bytes_to_skip = emit->bytecode_offset - emit->last_source_line_offset;
|
|
|
|
mp_uint_t lines_to_skip = source_line - emit->last_source_line;
|
2014-01-25 06:43:20 -05:00
|
|
|
emit_write_code_info_bytes_lines(emit, bytes_to_skip, lines_to_skip);
|
2014-05-10 05:36:38 -04:00
|
|
|
emit->last_source_line_offset = emit->bytecode_offset;
|
2014-01-18 18:24:36 -05:00
|
|
|
emit->last_source_line = source_line;
|
|
|
|
}
|
2020-02-26 23:36:53 -05:00
|
|
|
#else
|
2015-09-04 11:53:46 -04:00
|
|
|
(void)emit;
|
|
|
|
(void)source_line;
|
2020-02-26 23:36:53 -05:00
|
|
|
#endif
|
2014-01-18 18:24:36 -05:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_label_assign(emit_t *emit, mp_uint_t l) {
|
2023-03-07 22:10:02 -05:00
|
|
|
// Assigning a label ends any dead-code region, and all following opcodes
|
2022-06-17 09:14:32 -04:00
|
|
|
// should be emitted (until another unconditional flow control).
|
|
|
|
emit->suppress = false;
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
mp_emit_bc_adjust_stack_size(emit, 0);
|
2015-03-26 11:49:53 -04:00
|
|
|
if (emit->pass == MP_PASS_SCOPE) {
|
|
|
|
return;
|
|
|
|
}
|
2022-03-15 18:37:58 -04:00
|
|
|
|
|
|
|
// Label offsets can change from one pass to the next, but they must only
|
|
|
|
// decrease (ie code can only shrink). There will be multiple MP_PASS_EMIT
|
|
|
|
// stages until the labels no longer change, which is when the code size
|
|
|
|
// stays constant after a MP_PASS_EMIT.
|
2013-10-05 08:37:10 -04:00
|
|
|
assert(l < emit->max_num_labels);
|
2022-03-15 18:37:58 -04:00
|
|
|
assert(emit->pass == MP_PASS_STACK_SIZE || emit->bytecode_offset <= emit->label_offsets[l]);
|
|
|
|
|
|
|
|
// Assign label offset.
|
|
|
|
emit->label_offsets[l] = emit->bytecode_offset;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-22 07:58:25 -04:00
|
|
|
void mp_emit_bc_import(emit_t *emit, qstr qst, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_IMPORT_NAME + MP_EMIT_IMPORT_NAME == MP_BC_IMPORT_NAME);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_IMPORT_NAME + MP_EMIT_IMPORT_FROM == MP_BC_IMPORT_FROM);
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = kind == MP_EMIT_IMPORT_FROM ? 1 : -1;
|
2018-05-22 07:58:25 -04:00
|
|
|
if (kind == MP_EMIT_IMPORT_STAR) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, stack_adj, MP_BC_IMPORT_STAR);
|
2018-05-22 07:58:25 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, stack_adj, MP_BC_IMPORT_NAME + kind, qst);
|
2018-05-22 07:58:25 -04:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_load_const_tok(emit_t *emit, mp_token_kind_t tok) {
|
2019-08-22 00:36:27 -04:00
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_CONST_FALSE + (MP_TOKEN_KW_NONE - MP_TOKEN_KW_FALSE) == MP_BC_LOAD_CONST_NONE);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_CONST_FALSE + (MP_TOKEN_KW_TRUE - MP_TOKEN_KW_FALSE) == MP_BC_LOAD_CONST_TRUE);
|
|
|
|
if (tok == MP_TOKEN_ELLIPSIS) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_obj(emit, 1, MP_BC_LOAD_CONST_OBJ, MP_OBJ_FROM_PTR(&mp_const_ellipsis_obj));
|
2019-08-22 00:36:27 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 1, MP_BC_LOAD_CONST_FALSE + (tok - MP_TOKEN_KW_FALSE));
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_load_const_small_int(emit_t *emit, mp_int_t arg) {
|
2022-03-31 08:59:10 -04:00
|
|
|
assert(MP_SMALL_INT_FITS(arg));
|
2019-09-11 22:19:49 -04:00
|
|
|
if (-MP_BC_LOAD_CONST_SMALL_INT_MULTI_EXCESS <= arg
|
|
|
|
&& arg < MP_BC_LOAD_CONST_SMALL_INT_MULTI_NUM - MP_BC_LOAD_CONST_SMALL_INT_MULTI_EXCESS) {
|
|
|
|
emit_write_bytecode_byte(emit, 1,
|
|
|
|
MP_BC_LOAD_CONST_SMALL_INT_MULTI + MP_BC_LOAD_CONST_SMALL_INT_MULTI_EXCESS + arg);
|
py: Compress load-int, load-fast, store-fast, unop, binop bytecodes.
There is a lot potential in compress bytecodes and make more use of the
coding space. This patch introduces "multi" bytecodes which have their
argument included in the bytecode (by addition).
UNARY_OP and BINARY_OP now no longer take a 1 byte argument for the
opcode. Rather, the opcode is included in the first byte itself.
LOAD_FAST_[0,1,2] and STORE_FAST_[0,1,2] are removed in favour of their
multi versions, which can take an argument between 0 and 15 inclusive.
The majority of LOAD_FAST/STORE_FAST codes fit in this range and so this
saves a byte for each of these.
LOAD_CONST_SMALL_INT_MULTI is used to load small ints between -16 and 47
inclusive. Such ints are quite common and now only need 1 byte to
store, and now have much faster decoding.
In all this patch saves about 2% RAM for typically bytecode (1.8% on
64-bit test, 2.5% on pyboard test). It also reduces the binary size
(because bytecodes are simplified) and doesn't harm performance.
2014-10-25 11:43:46 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_int(emit, 1, MP_BC_LOAD_CONST_SMALL_INT, arg);
|
py: Compress load-int, load-fast, store-fast, unop, binop bytecodes.
There is a lot potential in compress bytecodes and make more use of the
coding space. This patch introduces "multi" bytecodes which have their
argument included in the bytecode (by addition).
UNARY_OP and BINARY_OP now no longer take a 1 byte argument for the
opcode. Rather, the opcode is included in the first byte itself.
LOAD_FAST_[0,1,2] and STORE_FAST_[0,1,2] are removed in favour of their
multi versions, which can take an argument between 0 and 15 inclusive.
The majority of LOAD_FAST/STORE_FAST codes fit in this range and so this
saves a byte for each of these.
LOAD_CONST_SMALL_INT_MULTI is used to load small ints between -16 and 47
inclusive. Such ints are quite common and now only need 1 byte to
store, and now have much faster decoding.
In all this patch saves about 2% RAM for typically bytecode (1.8% on
64-bit test, 2.5% on pyboard test). It also reduces the binary size
(because bytecodes are simplified) and doesn't harm performance.
2014-10-25 11:43:46 -04:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-06-25 10:42:13 -04:00
|
|
|
void mp_emit_bc_load_const_str(emit_t *emit, qstr qst) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, 1, MP_BC_LOAD_CONST_STRING, qst);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-11-27 07:41:25 -05:00
|
|
|
void mp_emit_bc_load_const_obj(emit_t *emit, mp_obj_t obj) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_obj(emit, 1, MP_BC_LOAD_CONST_OBJ, obj);
|
2015-01-13 10:55:54 -05:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_load_null(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 1, MP_BC_LOAD_NULL);
|
2017-09-13 06:36:06 -04:00
|
|
|
}
|
2014-03-31 06:59:23 -04:00
|
|
|
|
2018-05-18 10:11:04 -04:00
|
|
|
void mp_emit_bc_load_local(emit_t *emit, qstr qst, mp_uint_t local_num, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_FAST_N + MP_EMIT_IDOP_LOCAL_FAST == MP_BC_LOAD_FAST_N);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_FAST_N + MP_EMIT_IDOP_LOCAL_DEREF == MP_BC_LOAD_DEREF);
|
2015-01-20 07:47:20 -05:00
|
|
|
(void)qst;
|
2018-05-18 10:11:04 -04:00
|
|
|
if (kind == MP_EMIT_IDOP_LOCAL_FAST && local_num <= 15) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 1, MP_BC_LOAD_FAST_MULTI + local_num);
|
py: Compress load-int, load-fast, store-fast, unop, binop bytecodes.
There is a lot potential in compress bytecodes and make more use of the
coding space. This patch introduces "multi" bytecodes which have their
argument included in the bytecode (by addition).
UNARY_OP and BINARY_OP now no longer take a 1 byte argument for the
opcode. Rather, the opcode is included in the first byte itself.
LOAD_FAST_[0,1,2] and STORE_FAST_[0,1,2] are removed in favour of their
multi versions, which can take an argument between 0 and 15 inclusive.
The majority of LOAD_FAST/STORE_FAST codes fit in this range and so this
saves a byte for each of these.
LOAD_CONST_SMALL_INT_MULTI is used to load small ints between -16 and 47
inclusive. Such ints are quite common and now only need 1 byte to
store, and now have much faster decoding.
In all this patch saves about 2% RAM for typically bytecode (1.8% on
64-bit test, 2.5% on pyboard test). It also reduces the binary size
(because bytecodes are simplified) and doesn't harm performance.
2014-10-25 11:43:46 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, 1, MP_BC_LOAD_FAST_N + kind, local_num);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-22 07:16:30 -04:00
|
|
|
void mp_emit_bc_load_global(emit_t *emit, qstr qst, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_NAME + MP_EMIT_IDOP_GLOBAL_NAME == MP_BC_LOAD_NAME);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_LOAD_NAME + MP_EMIT_IDOP_GLOBAL_GLOBAL == MP_BC_LOAD_GLOBAL);
|
2015-01-20 07:47:20 -05:00
|
|
|
(void)qst;
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, 1, MP_BC_LOAD_NAME + kind, qst);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2017-04-18 19:45:59 -04:00
|
|
|
void mp_emit_bc_load_method(emit_t *emit, qstr qst, bool is_super) {
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = 1 - 2 * is_super;
|
|
|
|
emit_write_bytecode_byte_qstr(emit, stack_adj, is_super ? MP_BC_LOAD_SUPER_METHOD : MP_BC_LOAD_METHOD, qst);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_load_build_class(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 1, MP_BC_LOAD_BUILD_CLASS);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-22 07:31:56 -04:00
|
|
|
void mp_emit_bc_subscr(emit_t *emit, int kind) {
|
|
|
|
if (kind == MP_EMIT_SUBSCR_LOAD) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_LOAD_SUBSCR);
|
2018-05-22 07:31:56 -04:00
|
|
|
} else {
|
|
|
|
if (kind == MP_EMIT_SUBSCR_DELETE) {
|
|
|
|
mp_emit_bc_load_null(emit);
|
|
|
|
mp_emit_bc_rot_three(emit);
|
|
|
|
}
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -3, MP_BC_STORE_SUBSCR);
|
2018-05-22 07:31:56 -04:00
|
|
|
}
|
2014-04-17 17:10:53 -04:00
|
|
|
}
|
|
|
|
|
2018-05-22 07:43:41 -04:00
|
|
|
void mp_emit_bc_attr(emit_t *emit, qstr qst, int kind) {
|
|
|
|
if (kind == MP_EMIT_ATTR_LOAD) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, 0, MP_BC_LOAD_ATTR, qst);
|
2018-05-22 07:43:41 -04:00
|
|
|
} else {
|
|
|
|
if (kind == MP_EMIT_ATTR_DELETE) {
|
|
|
|
mp_emit_bc_load_null(emit);
|
|
|
|
mp_emit_bc_rot_two(emit);
|
|
|
|
}
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, -2, MP_BC_STORE_ATTR, qst);
|
2018-05-22 07:43:41 -04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-18 10:11:04 -04:00
|
|
|
void mp_emit_bc_store_local(emit_t *emit, qstr qst, mp_uint_t local_num, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_STORE_FAST_N + MP_EMIT_IDOP_LOCAL_FAST == MP_BC_STORE_FAST_N);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_STORE_FAST_N + MP_EMIT_IDOP_LOCAL_DEREF == MP_BC_STORE_DEREF);
|
2015-01-20 07:47:20 -05:00
|
|
|
(void)qst;
|
2018-05-18 10:11:04 -04:00
|
|
|
if (kind == MP_EMIT_IDOP_LOCAL_FAST && local_num <= 15) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_STORE_FAST_MULTI + local_num);
|
py: Compress load-int, load-fast, store-fast, unop, binop bytecodes.
There is a lot potential in compress bytecodes and make more use of the
coding space. This patch introduces "multi" bytecodes which have their
argument included in the bytecode (by addition).
UNARY_OP and BINARY_OP now no longer take a 1 byte argument for the
opcode. Rather, the opcode is included in the first byte itself.
LOAD_FAST_[0,1,2] and STORE_FAST_[0,1,2] are removed in favour of their
multi versions, which can take an argument between 0 and 15 inclusive.
The majority of LOAD_FAST/STORE_FAST codes fit in this range and so this
saves a byte for each of these.
LOAD_CONST_SMALL_INT_MULTI is used to load small ints between -16 and 47
inclusive. Such ints are quite common and now only need 1 byte to
store, and now have much faster decoding.
In all this patch saves about 2% RAM for typically bytecode (1.8% on
64-bit test, 2.5% on pyboard test). It also reduces the binary size
(because bytecodes are simplified) and doesn't harm performance.
2014-10-25 11:43:46 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, -1, MP_BC_STORE_FAST_N + kind, local_num);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-22 07:16:30 -04:00
|
|
|
void mp_emit_bc_store_global(emit_t *emit, qstr qst, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_STORE_NAME + MP_EMIT_IDOP_GLOBAL_NAME == MP_BC_STORE_NAME);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_STORE_NAME + MP_EMIT_IDOP_GLOBAL_GLOBAL == MP_BC_STORE_GLOBAL);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, -1, MP_BC_STORE_NAME + kind, qst);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-18 10:11:04 -04:00
|
|
|
void mp_emit_bc_delete_local(emit_t *emit, qstr qst, mp_uint_t local_num, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_DELETE_FAST + MP_EMIT_IDOP_LOCAL_FAST == MP_BC_DELETE_FAST);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_DELETE_FAST + MP_EMIT_IDOP_LOCAL_DEREF == MP_BC_DELETE_DEREF);
|
2015-01-20 07:47:20 -05:00
|
|
|
(void)qst;
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, 0, MP_BC_DELETE_FAST + kind, local_num);
|
2013-12-10 19:41:43 -05:00
|
|
|
}
|
|
|
|
|
2018-05-22 07:16:30 -04:00
|
|
|
void mp_emit_bc_delete_global(emit_t *emit, qstr qst, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_DELETE_NAME + MP_EMIT_IDOP_GLOBAL_NAME == MP_BC_DELETE_NAME);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_DELETE_NAME + MP_EMIT_IDOP_GLOBAL_GLOBAL == MP_BC_DELETE_GLOBAL);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_qstr(emit, 0, MP_BC_DELETE_NAME + kind, qst);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_dup_top(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 1, MP_BC_DUP_TOP);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_dup_top_two(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 2, MP_BC_DUP_TOP_TWO);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_pop_top(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_POP_TOP);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_rot_two(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 0, MP_BC_ROT_TWO);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_rot_three(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 0, MP_BC_ROT_THREE);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_jump(emit_t *emit, mp_uint_t label) {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, 0, MP_BC_JUMP, label);
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = true;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_pop_jump_if(emit_t *emit, bool cond, mp_uint_t label) {
|
2015-02-28 10:04:06 -05:00
|
|
|
if (cond) {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, -1, MP_BC_POP_JUMP_IF_TRUE, label);
|
2015-02-28 10:04:06 -05:00
|
|
|
} else {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, -1, MP_BC_POP_JUMP_IF_FALSE, label);
|
2015-02-28 10:04:06 -05:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_jump_if_or_pop(emit_t *emit, bool cond, mp_uint_t label) {
|
2015-02-28 10:04:06 -05:00
|
|
|
if (cond) {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, -1, MP_BC_JUMP_IF_TRUE_OR_POP, label);
|
2015-02-28 10:04:06 -05:00
|
|
|
} else {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, -1, MP_BC_JUMP_IF_FALSE_OR_POP, label);
|
2015-02-28 10:04:06 -05:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_unwind_jump(emit_t *emit, mp_uint_t label, mp_uint_t except_depth) {
|
2014-02-01 15:08:18 -05:00
|
|
|
if (except_depth == 0) {
|
2014-05-30 10:20:41 -04:00
|
|
|
if (label & MP_EMIT_BREAK_FROM_FOR) {
|
|
|
|
// need to pop the iterator if we are breaking out of a for loop
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_raw_byte(emit, MP_BC_POP_TOP);
|
2016-01-09 18:59:52 -05:00
|
|
|
// also pop the iter_buf
|
2017-03-23 01:36:08 -04:00
|
|
|
for (size_t i = 0; i < MP_OBJ_ITER_BUF_NSLOTS - 1; ++i) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_raw_byte(emit, MP_BC_POP_TOP);
|
2016-01-09 18:59:52 -05:00
|
|
|
}
|
2014-05-30 10:20:41 -04:00
|
|
|
}
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, 0, MP_BC_JUMP, label & ~MP_EMIT_BREAK_FROM_FOR);
|
2014-05-30 10:20:41 -04:00
|
|
|
} else {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, 0, MP_BC_UNWIND_JUMP, label & ~MP_EMIT_BREAK_FROM_FOR);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_raw_byte(emit, ((label & MP_EMIT_BREAK_FROM_FOR) ? 0x80 : 0) | except_depth);
|
2014-02-01 15:08:18 -05:00
|
|
|
}
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = true;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-22 08:33:26 -04:00
|
|
|
void mp_emit_bc_setup_block(emit_t *emit, mp_uint_t label, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_SETUP_WITH + MP_EMIT_SETUP_BLOCK_WITH == MP_BC_SETUP_WITH);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_SETUP_WITH + MP_EMIT_SETUP_BLOCK_EXCEPT == MP_BC_SETUP_EXCEPT);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_SETUP_WITH + MP_EMIT_SETUP_BLOCK_FINALLY == MP_BC_SETUP_FINALLY);
|
2016-09-26 22:37:21 -04:00
|
|
|
// The SETUP_WITH opcode pops ctx_mgr from the top of the stack
|
|
|
|
// and then pushes 3 entries: __exit__, ctx_mgr, as_value.
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = kind == MP_EMIT_SETUP_BLOCK_WITH ? 2 : 0;
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, stack_adj, MP_BC_SETUP_WITH + kind, label);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2016-04-07 03:50:38 -04:00
|
|
|
void mp_emit_bc_with_cleanup(emit_t *emit, mp_uint_t label) {
|
|
|
|
mp_emit_bc_load_const_tok(emit, MP_TOKEN_KW_NONE);
|
|
|
|
mp_emit_bc_label_assign(emit, label);
|
2019-08-22 00:55:16 -04:00
|
|
|
// The +2 is to ensure we have enough stack space to call the __exit__ method
|
|
|
|
emit_write_bytecode_byte(emit, 2, MP_BC_WITH_CLEANUP);
|
|
|
|
// Cancel the +2 above, plus the +2 from mp_emit_bc_setup_block(MP_EMIT_SETUP_BLOCK_WITH)
|
|
|
|
mp_emit_bc_adjust_stack_size(emit, -4);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_end_finally(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_END_FINALLY);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2016-01-09 18:59:52 -05:00
|
|
|
void mp_emit_bc_get_iter(emit_t *emit, bool use_stack) {
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = use_stack ? MP_OBJ_ITER_BUF_NSLOTS - 1 : 0;
|
|
|
|
emit_write_bytecode_byte(emit, stack_adj, use_stack ? MP_BC_GET_ITER_STACK : MP_BC_GET_ITER);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_for_iter(emit_t *emit, mp_uint_t label) {
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, 1, MP_BC_FOR_ITER, label);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2017-01-16 23:30:18 -05:00
|
|
|
void mp_emit_bc_for_iter_end(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
mp_emit_bc_adjust_stack_size(emit, -MP_OBJ_ITER_BUF_NSLOTS);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2019-02-14 20:18:59 -05:00
|
|
|
void mp_emit_bc_pop_except_jump(emit_t *emit, mp_uint_t label, bool within_exc_handler) {
|
|
|
|
(void)within_exc_handler;
|
2022-03-15 18:37:58 -04:00
|
|
|
emit_write_bytecode_byte_label(emit, 0, MP_BC_POP_EXCEPT_JUMP, label);
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = true;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_unary_op(emit_t *emit, mp_unary_op_t op) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 0, MP_BC_UNARY_OP_MULTI + op);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_binary_op(emit_t *emit, mp_binary_op_t op) {
|
2014-02-01 18:04:09 -05:00
|
|
|
bool invert = false;
|
2014-03-30 08:35:08 -04:00
|
|
|
if (op == MP_BINARY_OP_NOT_IN) {
|
2014-02-01 18:04:09 -05:00
|
|
|
invert = true;
|
2014-03-30 08:35:08 -04:00
|
|
|
op = MP_BINARY_OP_IN;
|
|
|
|
} else if (op == MP_BINARY_OP_IS_NOT) {
|
2014-02-01 18:04:09 -05:00
|
|
|
invert = true;
|
2014-03-30 08:35:08 -04:00
|
|
|
op = MP_BINARY_OP_IS;
|
2014-02-01 18:04:09 -05:00
|
|
|
}
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_BINARY_OP_MULTI + op);
|
2014-02-01 18:04:09 -05:00
|
|
|
if (invert) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, 0, MP_BC_UNARY_OP_MULTI + MP_UNARY_OP_NOT);
|
2014-02-01 18:04:09 -05:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-18 10:41:40 -04:00
|
|
|
void mp_emit_bc_build(emit_t *emit, mp_uint_t n_args, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_BUILD_TUPLE + MP_EMIT_BUILD_TUPLE == MP_BC_BUILD_TUPLE);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_BUILD_TUPLE + MP_EMIT_BUILD_LIST == MP_BC_BUILD_LIST);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_BUILD_TUPLE + MP_EMIT_BUILD_MAP == MP_BC_BUILD_MAP);
|
2018-05-22 08:18:42 -04:00
|
|
|
MP_STATIC_ASSERT(MP_BC_BUILD_TUPLE + MP_EMIT_BUILD_SET == MP_BC_BUILD_SET);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_BUILD_TUPLE + MP_EMIT_BUILD_SLICE == MP_BC_BUILD_SLICE);
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = kind == MP_EMIT_BUILD_MAP ? 1 : 1 - n_args;
|
|
|
|
emit_write_bytecode_byte_uint(emit, stack_adj, MP_BC_BUILD_TUPLE + kind, n_args);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_store_map(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -2, MP_BC_STORE_MAP);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2016-09-18 09:59:47 -04:00
|
|
|
void mp_emit_bc_store_comp(emit_t *emit, scope_kind_t kind, mp_uint_t collection_stack_index) {
|
2016-09-18 18:46:01 -04:00
|
|
|
int t;
|
2016-09-18 09:59:47 -04:00
|
|
|
int n;
|
|
|
|
if (kind == SCOPE_LIST_COMP) {
|
2016-09-18 18:46:01 -04:00
|
|
|
n = 0;
|
|
|
|
t = 0;
|
|
|
|
} else if (!MICROPY_PY_BUILTINS_SET || kind == SCOPE_DICT_COMP) {
|
|
|
|
n = 1;
|
|
|
|
t = 1;
|
|
|
|
} else if (MICROPY_PY_BUILTINS_SET) {
|
|
|
|
n = 0;
|
|
|
|
t = 2;
|
|
|
|
}
|
|
|
|
// the lower 2 bits of the opcode argument indicate the collection type
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, -1 - n, MP_BC_STORE_COMP, ((collection_stack_index + n) << 2) | t);
|
2016-09-18 09:59:47 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_unpack_sequence(emit_t *emit, mp_uint_t n_args) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, -1 + n_args, MP_BC_UNPACK_SEQUENCE, n_args);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_unpack_ex(emit_t *emit, mp_uint_t n_left, mp_uint_t n_right) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, -1 + n_left + n_right + 1, MP_BC_UNPACK_EX, n_left | (n_right << 8));
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_make_function(emit_t *emit, scope_t *scope, mp_uint_t n_pos_defaults, mp_uint_t n_kw_defaults) {
|
2014-03-31 10:18:37 -04:00
|
|
|
if (n_pos_defaults == 0 && n_kw_defaults == 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 07:22:47 -04:00
|
|
|
emit_write_bytecode_byte_child(emit, 1, MP_BC_MAKE_FUNCTION, scope->raw_code);
|
2014-02-01 13:29:40 -05:00
|
|
|
} else {
|
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 07:22:47 -04:00
|
|
|
emit_write_bytecode_byte_child(emit, -1, MP_BC_MAKE_FUNCTION_DEFARGS, scope->raw_code);
|
2014-02-01 08:05:04 -05:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_make_closure(emit_t *emit, scope_t *scope, mp_uint_t n_closed_over, mp_uint_t n_pos_defaults, mp_uint_t n_kw_defaults) {
|
2014-03-31 10:18:37 -04:00
|
|
|
if (n_pos_defaults == 0 && n_kw_defaults == 0) {
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = -n_closed_over + 1;
|
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 07:22:47 -04:00
|
|
|
emit_write_bytecode_byte_child(emit, stack_adj, MP_BC_MAKE_CLOSURE, scope->raw_code);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_raw_byte(emit, n_closed_over);
|
2014-03-26 17:14:59 -04:00
|
|
|
} else {
|
2014-04-20 12:50:40 -04:00
|
|
|
assert(n_closed_over <= 255);
|
2019-08-22 00:55:16 -04:00
|
|
|
int stack_adj = -2 - (mp_int_t)n_closed_over + 1;
|
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 07:22:47 -04:00
|
|
|
emit_write_bytecode_byte_child(emit, stack_adj, MP_BC_MAKE_CLOSURE_DEFARGS, scope->raw_code);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_raw_byte(emit, n_closed_over);
|
2014-03-26 17:14:59 -04:00
|
|
|
}
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2019-08-22 00:55:16 -04:00
|
|
|
STATIC void emit_bc_call_function_method_helper(emit_t *emit, int stack_adj, mp_uint_t bytecode_base, mp_uint_t n_positional, mp_uint_t n_keyword, mp_uint_t star_flags) {
|
2014-04-09 07:43:17 -04:00
|
|
|
if (star_flags) {
|
2020-03-24 22:39:46 -04:00
|
|
|
// each positional arg is one object, each kwarg is two objects, the key
|
|
|
|
// and the value and one extra object for the star args bitmap.
|
|
|
|
stack_adj -= (int)n_positional + 2 * (int)n_keyword + 1;
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte_uint(emit, stack_adj, bytecode_base + 1, (n_keyword << 8) | n_positional); // TODO make it 2 separate uints?
|
2013-10-04 14:53:11 -04:00
|
|
|
} else {
|
2019-08-22 00:55:16 -04:00
|
|
|
stack_adj -= (int)n_positional + 2 * (int)n_keyword;
|
|
|
|
emit_write_bytecode_byte_uint(emit, stack_adj, bytecode_base, (n_keyword << 8) | n_positional); // TODO make it 2 separate uints?
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
2014-03-31 06:59:23 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_call_function(emit_t *emit, mp_uint_t n_positional, mp_uint_t n_keyword, mp_uint_t star_flags) {
|
2014-04-09 07:43:17 -04:00
|
|
|
emit_bc_call_function_method_helper(emit, 0, MP_BC_CALL_FUNCTION, n_positional, n_keyword, star_flags);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_call_method(emit_t *emit, mp_uint_t n_positional, mp_uint_t n_keyword, mp_uint_t star_flags) {
|
2014-04-09 07:43:17 -04:00
|
|
|
emit_bc_call_function_method_helper(emit, -1, MP_BC_CALL_METHOD, n_positional, n_keyword, star_flags);
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_return_value(emit_t *emit) {
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -1, MP_BC_RETURN_VALUE);
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = true;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_raise_varargs(emit_t *emit, mp_uint_t n_args) {
|
2019-08-21 22:39:07 -04:00
|
|
|
MP_STATIC_ASSERT(MP_BC_RAISE_LAST + 1 == MP_BC_RAISE_OBJ);
|
|
|
|
MP_STATIC_ASSERT(MP_BC_RAISE_LAST + 2 == MP_BC_RAISE_FROM);
|
2016-10-25 05:05:33 -04:00
|
|
|
assert(n_args <= 2);
|
2019-08-21 22:39:07 -04:00
|
|
|
emit_write_bytecode_byte(emit, -n_args, MP_BC_RAISE_LAST + n_args);
|
2022-06-17 09:14:32 -04:00
|
|
|
emit->suppress = true;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2018-05-18 10:30:42 -04:00
|
|
|
void mp_emit_bc_yield(emit_t *emit, int kind) {
|
|
|
|
MP_STATIC_ASSERT(MP_BC_YIELD_VALUE + 1 == MP_BC_YIELD_FROM);
|
2019-08-22 00:55:16 -04:00
|
|
|
emit_write_bytecode_byte(emit, -kind, MP_BC_YIELD_VALUE + kind);
|
2014-05-07 12:24:22 -04:00
|
|
|
emit->scope->scope_flags |= MP_SCOPE_FLAG_GENERATOR;
|
2013-10-04 14:53:11 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_start_except_handler(emit_t *emit) {
|
2016-09-26 22:37:21 -04:00
|
|
|
mp_emit_bc_adjust_stack_size(emit, 4); // stack adjust for the exception instance, +3 for possible UNWIND_JUMP state
|
2014-06-30 00:17:25 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
void mp_emit_bc_end_except_handler(emit_t *emit) {
|
2016-09-26 22:37:21 -04:00
|
|
|
mp_emit_bc_adjust_stack_size(emit, -3); // stack adjust
|
2014-06-30 00:17:25 -04:00
|
|
|
}
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
#if MICROPY_EMIT_NATIVE
|
2013-10-05 13:08:26 -04:00
|
|
|
const emit_method_table_t emit_bc_method_table = {
|
2019-03-08 18:59:25 -05:00
|
|
|
#if MICROPY_DYNAMIC_COMPILER
|
|
|
|
NULL,
|
|
|
|
NULL,
|
|
|
|
#endif
|
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_start_pass,
|
|
|
|
mp_emit_bc_end_pass,
|
|
|
|
mp_emit_bc_adjust_stack_size,
|
|
|
|
mp_emit_bc_set_source_line,
|
2013-10-05 07:19:06 -04:00
|
|
|
|
2015-03-26 10:42:40 -04:00
|
|
|
{
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_load_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_load_global,
|
2015-03-26 10:42:40 -04:00
|
|
|
},
|
|
|
|
{
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_store_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_store_global,
|
2015-03-26 10:42:40 -04:00
|
|
|
},
|
|
|
|
{
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_delete_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_delete_global,
|
2015-03-26 10:42:40 -04:00
|
|
|
},
|
2013-10-05 09:17:09 -04:00
|
|
|
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_label_assign,
|
2018-05-22 07:58:25 -04:00
|
|
|
mp_emit_bc_import,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_load_const_tok,
|
|
|
|
mp_emit_bc_load_const_small_int,
|
|
|
|
mp_emit_bc_load_const_str,
|
|
|
|
mp_emit_bc_load_const_obj,
|
|
|
|
mp_emit_bc_load_null,
|
|
|
|
mp_emit_bc_load_method,
|
|
|
|
mp_emit_bc_load_build_class,
|
2018-05-22 07:31:56 -04:00
|
|
|
mp_emit_bc_subscr,
|
2018-05-22 07:43:41 -04:00
|
|
|
mp_emit_bc_attr,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_dup_top,
|
|
|
|
mp_emit_bc_dup_top_two,
|
|
|
|
mp_emit_bc_pop_top,
|
|
|
|
mp_emit_bc_rot_two,
|
|
|
|
mp_emit_bc_rot_three,
|
|
|
|
mp_emit_bc_jump,
|
|
|
|
mp_emit_bc_pop_jump_if,
|
|
|
|
mp_emit_bc_jump_if_or_pop,
|
|
|
|
mp_emit_bc_unwind_jump,
|
2018-05-22 08:33:26 -04:00
|
|
|
mp_emit_bc_setup_block,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_with_cleanup,
|
|
|
|
mp_emit_bc_end_finally,
|
|
|
|
mp_emit_bc_get_iter,
|
|
|
|
mp_emit_bc_for_iter,
|
|
|
|
mp_emit_bc_for_iter_end,
|
2019-02-14 20:18:59 -05:00
|
|
|
mp_emit_bc_pop_except_jump,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_unary_op,
|
|
|
|
mp_emit_bc_binary_op,
|
2018-05-18 10:41:40 -04:00
|
|
|
mp_emit_bc_build,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_store_map,
|
2016-09-18 09:59:47 -04:00
|
|
|
mp_emit_bc_store_comp,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_unpack_sequence,
|
|
|
|
mp_emit_bc_unpack_ex,
|
|
|
|
mp_emit_bc_make_function,
|
|
|
|
mp_emit_bc_make_closure,
|
|
|
|
mp_emit_bc_call_function,
|
|
|
|
mp_emit_bc_call_method,
|
|
|
|
mp_emit_bc_return_value,
|
|
|
|
mp_emit_bc_raise_varargs,
|
2018-05-18 10:30:42 -04:00
|
|
|
mp_emit_bc_yield,
|
2015-03-26 12:44:14 -04:00
|
|
|
|
|
|
|
mp_emit_bc_start_except_handler,
|
|
|
|
mp_emit_bc_end_except_handler,
|
|
|
|
};
|
|
|
|
#else
|
|
|
|
const mp_emit_method_table_id_ops_t mp_emit_bc_method_table_load_id_ops = {
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_load_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_load_global,
|
|
|
|
};
|
|
|
|
|
|
|
|
const mp_emit_method_table_id_ops_t mp_emit_bc_method_table_store_id_ops = {
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_store_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_store_global,
|
|
|
|
};
|
|
|
|
|
|
|
|
const mp_emit_method_table_id_ops_t mp_emit_bc_method_table_delete_id_ops = {
|
2018-05-18 10:11:04 -04:00
|
|
|
mp_emit_bc_delete_local,
|
2015-03-26 12:44:14 -04:00
|
|
|
mp_emit_bc_delete_global,
|
2013-10-05 07:19:06 -04:00
|
|
|
};
|
2015-03-26 12:44:14 -04:00
|
|
|
#endif
|
2015-12-18 07:35:44 -05:00
|
|
|
|
2020-04-16 03:13:57 -04:00
|
|
|
#endif // MICROPY_ENABLE_COMPILER
|