micropython/py/vm.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1475 lines
62 KiB
C
Raw Normal View History

/*
* This file is part of the MicroPython project, http://micropython.org/
*
* The MIT License (MIT)
*
* Copyright (c) 2013-2019 Damien P. George
* Copyright (c) 2014-2015 Paul Sokolovsky
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
2013-10-04 19:53:11 +01:00
#include <stdio.h>
#include <string.h>
2013-10-04 19:53:11 +01:00
#include <assert.h>
#include "py/emitglue.h"
#include "py/objtype.h"
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#include "py/objfun.h"
#include "py/runtime.h"
#include "py/bc0.h"
#include "py/profile.h"
2013-10-04 19:53:11 +01:00
// *FORMAT-OFF*
#if 0
#if MICROPY_PY_THREAD
#define TRACE_PREFIX mp_printf(&mp_plat_print, "ts=%p sp=%d ", mp_thread_get_state(), (int)(sp - &code_state->state[0] + 1))
#else
#define TRACE_PREFIX mp_printf(&mp_plat_print, "sp=%d ", (int)(sp - &code_state->state[0] + 1))
#endif
#define TRACE(ip) TRACE_PREFIX; mp_bytecode_print2(&mp_plat_print, ip, 1, code_state->fun_bc->child_table, &code_state->fun_bc->context->constants);
#else
#define TRACE(ip)
#endif
2014-01-31 17:45:15 +00:00
// Value stack grows up (this makes it incompatible with native C stack, but
// makes sure that arguments to functions are in natural order arg1..argN
// (Python semantics mandates left-to-right evaluation order, including for
// function arguments). Stack pointer is pre-incremented and points at the
// top element.
// Exception stack also grows up, top element is also pointed at.
#define DECODE_UINT \
mp_uint_t unum = 0; \
do { \
unum = (unum << 7) + (*ip & 0x7f); \
} while ((*ip++ & 0x80) != 0)
#define DECODE_ULABEL \
size_t ulab; \
do { \
if (ip[0] & 0x80) { \
ulab = ((ip[0] & 0x7f) | (ip[1] << 7)); \
ip += 2; \
} else { \
ulab = ip[0]; \
ip += 1; \
} \
} while (0)
#define DECODE_SLABEL \
size_t slab; \
do { \
if (ip[0] & 0x80) { \
slab = ((ip[0] & 0x7f) | (ip[1] << 7)) - 0x4000; \
ip += 2; \
} else { \
slab = ip[0] - 0x40; \
ip += 1; \
} \
} while (0)
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#if MICROPY_EMIT_BYTECODE_USES_QSTR_TABLE
#define DECODE_QSTR \
DECODE_UINT; \
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
qstr qst = qstr_table[unum]
#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 12:22:47 +01:00
#define DECODE_QSTR \
DECODE_UINT; \
qstr qst = unum;
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#define DECODE_PTR \
DECODE_UINT; \
void *ptr = (void *)(uintptr_t)code_state->fun_bc->child_table[unum]
#define DECODE_OBJ \
DECODE_UINT; \
mp_obj_t obj = (mp_obj_t)code_state->fun_bc->context->constants.obj_table[unum]
#define PUSH(val) *++sp = (val)
#define POP() (*sp--)
2013-12-10 17:27:24 +00:00
#define TOP() (*sp)
#define SET_TOP(val) *sp = (val)
2013-10-04 19:53:11 +01:00
#if MICROPY_PY_SYS_EXC_INFO
#define CLEAR_SYS_EXC_INFO() MP_STATE_VM(cur_exception) = NULL;
#else
#define CLEAR_SYS_EXC_INFO()
#endif
#define PUSH_EXC_BLOCK(with_or_finally) do { \
DECODE_ULABEL; /* except labels are always forward */ \
++exc_sp; \
exc_sp->handler = ip + ulab; \
exc_sp->val_sp = MP_TAGPTR_MAKE(sp, ((with_or_finally) << 1)); \
exc_sp->prev_exc = NULL; \
} while (0)
#define POP_EXC_BLOCK() \
exc_sp--; /* pop back to previous exception handler */ \
CLEAR_SYS_EXC_INFO() /* just clear sys.exc_info(), not compliant, but it shouldn't be used in 1st place */
#define CANCEL_ACTIVE_FINALLY(sp) do { \
if (mp_obj_is_small_int(sp[-1])) { \
/* Stack: (..., prev_dest_ip, prev_cause, dest_ip) */ \
/* Cancel the unwind through the previous finally, replace with current one */ \
sp[-2] = sp[0]; \
sp -= 2; \
} else { \
assert(sp[-1] == mp_const_none || mp_obj_is_exception_instance(sp[-1])); \
/* Stack: (..., None/exception, dest_ip) */ \
/* Silence the finally's exception value (may be None or an exception) */ \
sp[-1] = sp[0]; \
--sp; \
} \
} while (0)
#if MICROPY_PY_SYS_SETTRACE
#define FRAME_SETUP() do { \
assert(code_state != code_state->prev_state); \
MP_STATE_THREAD(current_code_state) = code_state; \
assert(code_state != code_state->prev_state); \
} while(0)
#define FRAME_ENTER() do { \
assert(code_state != code_state->prev_state); \
code_state->prev_state = MP_STATE_THREAD(current_code_state); \
assert(code_state != code_state->prev_state); \
if (!mp_prof_is_executing) { \
mp_prof_frame_enter(code_state); \
} \
} while(0)
#define FRAME_LEAVE() do { \
assert(code_state != code_state->prev_state); \
MP_STATE_THREAD(current_code_state) = code_state->prev_state; \
assert(code_state != code_state->prev_state); \
} while(0)
#define FRAME_UPDATE() do { \
assert(MP_STATE_THREAD(current_code_state) == code_state); \
if (!mp_prof_is_executing) { \
code_state->frame = MP_OBJ_TO_PTR(mp_prof_frame_update(code_state)); \
} \
} while(0)
#define TRACE_TICK(current_ip, current_sp, is_exception) do { \
assert(code_state != code_state->prev_state); \
assert(MP_STATE_THREAD(current_code_state) == code_state); \
if (!mp_prof_is_executing && code_state->frame && MP_STATE_THREAD(prof_trace_callback)) { \
MP_PROF_INSTR_DEBUG_PRINT(code_state->ip); \
} \
if (!mp_prof_is_executing && code_state->frame && code_state->frame->callback) { \
mp_prof_instr_tick(code_state, is_exception); \
} \
} while(0)
#else // MICROPY_PY_SYS_SETTRACE
#define FRAME_SETUP()
#define FRAME_ENTER()
#define FRAME_LEAVE()
#define FRAME_UPDATE()
#define TRACE_TICK(current_ip, current_sp, is_exception)
#endif // MICROPY_PY_SYS_SETTRACE
// fastn has items in reverse order (fastn[0] is local[0], fastn[-1] is local[1], etc)
// sp points to bottom of stack which grows up
// returns:
// MP_VM_RETURN_NORMAL, sp valid, return value in *sp
// MP_VM_RETURN_YIELD, ip, sp valid, yielded value in *sp
// MP_VM_RETURN_EXCEPTION, exception in state[0]
mp_vm_return_kind_t MICROPY_WRAP_MP_EXECUTE_BYTECODE(mp_execute_bytecode)(mp_code_state_t *code_state, volatile mp_obj_t inject_exc) {
#define SELECTIVE_EXC_IP (0)
#if SELECTIVE_EXC_IP
#define MARK_EXC_IP_SELECTIVE() { code_state->ip = ip; } /* stores ip 1 byte past last opcode */
#define MARK_EXC_IP_GLOBAL()
#else
#define MARK_EXC_IP_SELECTIVE()
#define MARK_EXC_IP_GLOBAL() { code_state->ip = ip; } /* stores ip pointing to last opcode */
#endif
#if MICROPY_OPT_COMPUTED_GOTO
#include "py/vmentrytable.h"
#define DISPATCH() do { \
TRACE(ip); \
MARK_EXC_IP_GLOBAL(); \
TRACE_TICK(ip, sp, false); \
goto *entry_table[*ip++]; \
} while (0)
#define DISPATCH_WITH_PEND_EXC_CHECK() goto pending_exception_check
#define ENTRY(op) entry_##op
#define ENTRY_DEFAULT entry_default
#else
#define DISPATCH() goto dispatch_loop
#define DISPATCH_WITH_PEND_EXC_CHECK() goto pending_exception_check
#define ENTRY(op) case op
#define ENTRY_DEFAULT default
#endif
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
// nlr_raise needs to be implemented as a goto, so that the C compiler's flow analyser
// sees that it's possible for us to jump from the dispatch loop to the exception
// handler. Without this, the code may have a different stack layout in the dispatch
// loop and the exception handler, leading to very obscure bugs.
#define RAISE(o) do { nlr_pop(); nlr.ret_val = MP_OBJ_TO_PTR(o); goto exception_handler; } while (0)
2013-10-04 19:53:11 +01:00
#if MICROPY_STACKLESS
run_code_state: ;
#endif
FRAME_ENTER();
#if MICROPY_STACKLESS
run_code_state_from_return: ;
#endif
FRAME_SETUP();
// Pointers which are constant for particular invocation of mp_execute_bytecode()
mp_obj_t * /*const*/ fastn;
mp_exc_stack_t * /*const*/ exc_stack;
{
size_t n_state = code_state->n_state;
fastn = &code_state->state[n_state - 1];
exc_stack = (mp_exc_stack_t*)(code_state->state + n_state);
}
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
// variables that are visible to the exception handler (declared volatile)
mp_exc_stack_t *volatile exc_sp = MP_CODE_STATE_EXC_SP_IDX_TO_PTR(exc_stack, code_state->exc_sp_idx); // stack grows up, exc_sp points to top of stack
2013-10-15 23:46:01 +01:00
#if MICROPY_PY_THREAD_GIL && MICROPY_PY_THREAD_GIL_VM_DIVISOR
// This needs to be volatile and outside the VM loop so it persists across handling
// of any exceptions. Otherwise it's possible that the VM never gives up the GIL.
volatile int gil_divisor = MICROPY_PY_THREAD_GIL_VM_DIVISOR;
#endif
// outer exception handling loop
2013-10-04 19:53:11 +01:00
for (;;) {
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
nlr_buf_t nlr;
outer_dispatch_loop:
if (nlr_push(&nlr) == 0) {
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
// local variables that are not visible to the exception handler
const byte *ip = code_state->ip;
mp_obj_t *sp = code_state->sp;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#if MICROPY_EMIT_BYTECODE_USES_QSTR_TABLE
const qstr_short_t *qstr_table = code_state->fun_bc->context->constants.qstr_table;
#endif
mp_obj_t obj_shared;
MICROPY_VM_HOOK_INIT
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
// If we have exception to inject, now that we finish setting up
// execution context, raise it. This works as if MP_BC_RAISE_OBJ
// bytecode was executed.
2014-03-26 15:36:12 +00:00
// Injecting exc into yield from generator is a special case,
// handled by MP_BC_YIELD_FROM itself
if (inject_exc != MP_OBJ_NULL && *ip != MP_BC_YIELD_FROM) {
mp_obj_t exc = inject_exc;
inject_exc = MP_OBJ_NULL;
exc = mp_make_raise_obj(exc);
RAISE(exc);
}
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
// loop to execute byte code
for (;;) {
dispatch_loop:
#if MICROPY_OPT_COMPUTED_GOTO
DISPATCH();
#else
TRACE(ip);
MARK_EXC_IP_GLOBAL();
TRACE_TICK(ip, sp, false);
switch (*ip++) {
#endif
ENTRY(MP_BC_LOAD_CONST_FALSE):
PUSH(mp_const_false);
DISPATCH();
ENTRY(MP_BC_LOAD_CONST_NONE):
PUSH(mp_const_none);
DISPATCH();
ENTRY(MP_BC_LOAD_CONST_TRUE):
PUSH(mp_const_true);
DISPATCH();
ENTRY(MP_BC_LOAD_CONST_SMALL_INT): {
mp_uint_t num = 0;
if ((ip[0] & 0x40) != 0) {
// Number is negative
num--;
}
do {
num = (num << 7) | (*ip & 0x7f);
} while ((*ip++ & 0x80) != 0);
PUSH(MP_OBJ_NEW_SMALL_INT(num));
DISPATCH();
}
ENTRY(MP_BC_LOAD_CONST_STRING): {
DECODE_QSTR;
PUSH(MP_OBJ_NEW_QSTR(qst));
DISPATCH();
}
ENTRY(MP_BC_LOAD_CONST_OBJ): {
DECODE_OBJ;
PUSH(obj);
DISPATCH();
}
ENTRY(MP_BC_LOAD_NULL):
PUSH(MP_OBJ_NULL);
DISPATCH();
ENTRY(MP_BC_LOAD_FAST_N): {
DECODE_UINT;
obj_shared = fastn[-unum];
load_check:
if (obj_shared == MP_OBJ_NULL) {
local_name_error: {
MARK_EXC_IP_SELECTIVE();
mp_obj_t obj = mp_obj_new_exception_msg(&mp_type_NameError, MP_ERROR_TEXT("local variable referenced before assignment"));
RAISE(obj);
}
}
PUSH(obj_shared);
DISPATCH();
}
ENTRY(MP_BC_LOAD_DEREF): {
DECODE_UINT;
obj_shared = mp_obj_cell_get(fastn[-unum]);
goto load_check;
}
ENTRY(MP_BC_LOAD_NAME): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
PUSH(mp_load_name(qst));
DISPATCH();
}
ENTRY(MP_BC_LOAD_GLOBAL): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
PUSH(mp_load_global(qst));
DISPATCH();
}
ENTRY(MP_BC_LOAD_ATTR): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_obj_t top = TOP();
mp_obj_t obj;
#if MICROPY_OPT_LOAD_ATTR_FAST_PATH
// For the specific case of an instance type, it implements .attr
// and forwards to its members map. Attribute lookups on instance
// types are extremely common, so avoid all the other checks and
// calls that normally happen first.
mp_map_elem_t *elem = NULL;
if (mp_obj_is_instance_type(mp_obj_get_type(top))) {
mp_obj_instance_t *self = MP_OBJ_TO_PTR(top);
elem = mp_map_lookup(&self->members, MP_OBJ_NEW_QSTR(qst), MP_MAP_LOOKUP);
}
if (elem) {
obj = elem->value;
} else
#endif
{
obj = mp_load_attr(top, qst);
}
SET_TOP(obj);
DISPATCH();
}
ENTRY(MP_BC_LOAD_METHOD): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_load_method(*sp, qst, sp);
sp += 1;
DISPATCH();
}
ENTRY(MP_BC_LOAD_SUPER_METHOD): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
sp -= 1;
mp_load_super_method(qst, sp - 1);
DISPATCH();
}
ENTRY(MP_BC_LOAD_BUILD_CLASS):
MARK_EXC_IP_SELECTIVE();
PUSH(mp_load_build_class());
DISPATCH();
ENTRY(MP_BC_LOAD_SUBSCR): {
MARK_EXC_IP_SELECTIVE();
mp_obj_t index = POP();
SET_TOP(mp_obj_subscr(TOP(), index, MP_OBJ_SENTINEL));
DISPATCH();
}
ENTRY(MP_BC_STORE_FAST_N): {
DECODE_UINT;
fastn[-unum] = POP();
DISPATCH();
}
ENTRY(MP_BC_STORE_DEREF): {
DECODE_UINT;
mp_obj_cell_set(fastn[-unum], POP());
DISPATCH();
}
ENTRY(MP_BC_STORE_NAME): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_store_name(qst, POP());
DISPATCH();
}
ENTRY(MP_BC_STORE_GLOBAL): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_store_global(qst, POP());
DISPATCH();
}
ENTRY(MP_BC_STORE_ATTR): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_store_attr(sp[0], qst, sp[-1]);
sp -= 2;
DISPATCH();
}
ENTRY(MP_BC_STORE_SUBSCR):
MARK_EXC_IP_SELECTIVE();
mp_obj_subscr(sp[-1], sp[0], sp[-2]);
sp -= 3;
DISPATCH();
ENTRY(MP_BC_DELETE_FAST): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
if (fastn[-unum] == MP_OBJ_NULL) {
goto local_name_error;
}
fastn[-unum] = MP_OBJ_NULL;
DISPATCH();
}
ENTRY(MP_BC_DELETE_DEREF): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
if (mp_obj_cell_get(fastn[-unum]) == MP_OBJ_NULL) {
goto local_name_error;
}
mp_obj_cell_set(fastn[-unum], MP_OBJ_NULL);
DISPATCH();
}
ENTRY(MP_BC_DELETE_NAME): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_delete_name(qst);
DISPATCH();
}
ENTRY(MP_BC_DELETE_GLOBAL): {
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_delete_global(qst);
DISPATCH();
}
ENTRY(MP_BC_DUP_TOP): {
mp_obj_t top = TOP();
PUSH(top);
DISPATCH();
}
ENTRY(MP_BC_DUP_TOP_TWO):
sp += 2;
sp[0] = sp[-2];
sp[-1] = sp[-3];
DISPATCH();
ENTRY(MP_BC_POP_TOP):
sp -= 1;
DISPATCH();
ENTRY(MP_BC_ROT_TWO): {
mp_obj_t top = sp[0];
sp[0] = sp[-1];
sp[-1] = top;
DISPATCH();
}
ENTRY(MP_BC_ROT_THREE): {
mp_obj_t top = sp[0];
sp[0] = sp[-1];
sp[-1] = sp[-2];
sp[-2] = top;
DISPATCH();
}
ENTRY(MP_BC_JUMP): {
DECODE_SLABEL;
ip += slab;
DISPATCH_WITH_PEND_EXC_CHECK();
}
ENTRY(MP_BC_POP_JUMP_IF_TRUE): {
DECODE_SLABEL;
if (mp_obj_is_true(POP())) {
ip += slab;
}
DISPATCH_WITH_PEND_EXC_CHECK();
}
2013-11-09 20:12:32 +00:00
ENTRY(MP_BC_POP_JUMP_IF_FALSE): {
DECODE_SLABEL;
if (!mp_obj_is_true(POP())) {
ip += slab;
}
DISPATCH_WITH_PEND_EXC_CHECK();
}
2013-11-09 20:12:32 +00:00
ENTRY(MP_BC_JUMP_IF_TRUE_OR_POP): {
DECODE_ULABEL;
if (mp_obj_is_true(TOP())) {
ip += ulab;
} else {
sp--;
}
DISPATCH_WITH_PEND_EXC_CHECK();
}
ENTRY(MP_BC_JUMP_IF_FALSE_OR_POP): {
DECODE_ULABEL;
if (mp_obj_is_true(TOP())) {
sp--;
} else {
ip += ulab;
}
DISPATCH_WITH_PEND_EXC_CHECK();
}
ENTRY(MP_BC_SETUP_WITH): {
MARK_EXC_IP_SELECTIVE();
// stack: (..., ctx_mgr)
mp_obj_t obj = TOP();
mp_load_method(obj, MP_QSTR___exit__, sp);
mp_load_method(obj, MP_QSTR___enter__, sp + 2);
mp_obj_t ret = mp_call_method_n_kw(0, 0, sp + 2);
sp += 1;
PUSH_EXC_BLOCK(1);
PUSH(ret);
// stack: (..., __exit__, ctx_mgr, as_value)
DISPATCH();
}
ENTRY(MP_BC_WITH_CLEANUP): {
MARK_EXC_IP_SELECTIVE();
// Arriving here, there's "exception control block" on top of stack,
// and __exit__ method (with self) underneath it. Bytecode calls __exit__,
// and "deletes" it off stack, shifting "exception control block"
// to its place.
// The bytecode emitter ensures that there is enough space on the Python
// value stack to hold the __exit__ method plus an additional 4 entries.
if (TOP() == mp_const_none) {
// stack: (..., __exit__, ctx_mgr, None)
sp[1] = mp_const_none;
sp[2] = mp_const_none;
sp -= 2;
mp_call_method_n_kw(3, 0, sp);
SET_TOP(mp_const_none);
} else if (mp_obj_is_small_int(TOP())) {
// Getting here there are two distinct cases:
// - unwind return, stack: (..., __exit__, ctx_mgr, ret_val, SMALL_INT(-1))
// - unwind jump, stack: (..., __exit__, ctx_mgr, dest_ip, SMALL_INT(num_exc))
// For both cases we do exactly the same thing.
mp_obj_t data = sp[-1];
mp_obj_t cause = sp[0];
sp[-1] = mp_const_none;
sp[0] = mp_const_none;
sp[1] = mp_const_none;
mp_call_method_n_kw(3, 0, sp - 3);
sp[-3] = data;
sp[-2] = cause;
sp -= 2; // we removed (__exit__, ctx_mgr)
} else {
assert(mp_obj_is_exception_instance(TOP()));
// stack: (..., __exit__, ctx_mgr, exc_instance)
// Need to pass (exc_type, exc_instance, None) as arguments to __exit__.
sp[1] = sp[0];
sp[0] = MP_OBJ_FROM_PTR(mp_obj_get_type(sp[0]));
sp[2] = mp_const_none;
sp -= 2;
mp_obj_t ret_value = mp_call_method_n_kw(3, 0, sp);
if (mp_obj_is_true(ret_value)) {
// We need to silence/swallow the exception. This is done
// by popping the exception and the __exit__ handler and
// replacing it with None, which signals END_FINALLY to just
// execute the finally handler normally.
SET_TOP(mp_const_none);
} else {
// We need to re-raise the exception. We pop __exit__ handler
// by copying the exception instance down to the new top-of-stack.
sp[0] = sp[3];
}
}
DISPATCH();
}
ENTRY(MP_BC_UNWIND_JUMP): {
MARK_EXC_IP_SELECTIVE();
DECODE_SLABEL;
PUSH((mp_obj_t)(mp_uint_t)(uintptr_t)(ip + slab)); // push destination ip for jump
PUSH((mp_obj_t)(mp_uint_t)(*ip)); // push number of exception handlers to unwind (0x80 bit set if we also need to pop stack)
unwind_jump:;
mp_uint_t unum = (mp_uint_t)POP(); // get number of exception handlers to unwind
while ((unum & 0x7f) > 0) {
unum -= 1;
assert(exc_sp >= exc_stack);
if (MP_TAGPTR_TAG1(exc_sp->val_sp)) {
if (exc_sp->handler > ip) {
// Found a finally handler that isn't active; run it.
// Getting here the stack looks like:
// (..., X, dest_ip)
// where X is pointed to by exc_sp->val_sp and in the case
// of a "with" block contains the context manager info.
assert(&sp[-1] == MP_TAGPTR_PTR(exc_sp->val_sp));
// We're going to run "finally" code as a coroutine
// (not calling it recursively). Set up a sentinel
// on the stack so it can return back to us when it is
// done (when WITH_CLEANUP or END_FINALLY reached).
// The sentinel is the number of exception handlers left to
// unwind, which is a non-negative integer.
PUSH(MP_OBJ_NEW_SMALL_INT(unum));
ip = exc_sp->handler;
goto dispatch_loop;
} else {
// Found a finally handler that is already active; cancel it.
CANCEL_ACTIVE_FINALLY(sp);
}
}
POP_EXC_BLOCK();
}
ip = (const byte*)MP_OBJ_TO_PTR(POP()); // pop destination ip for jump
if (unum != 0) {
// pop the exhausted iterator
sp -= MP_OBJ_ITER_BUF_NSLOTS;
}
DISPATCH_WITH_PEND_EXC_CHECK();
}
ENTRY(MP_BC_SETUP_EXCEPT):
ENTRY(MP_BC_SETUP_FINALLY): {
MARK_EXC_IP_SELECTIVE();
#if SELECTIVE_EXC_IP
PUSH_EXC_BLOCK((code_state->ip[-1] == MP_BC_SETUP_FINALLY) ? 1 : 0);
#else
PUSH_EXC_BLOCK((code_state->ip[0] == MP_BC_SETUP_FINALLY) ? 1 : 0);
#endif
DISPATCH();
}
ENTRY(MP_BC_END_FINALLY):
MARK_EXC_IP_SELECTIVE();
// if TOS is None, just pops it and continues
// if TOS is an integer, finishes coroutine and returns control to caller
// if TOS is an exception, reraises the exception
assert(exc_sp >= exc_stack);
POP_EXC_BLOCK();
if (TOP() == mp_const_none) {
sp--;
} else if (mp_obj_is_small_int(TOP())) {
// We finished "finally" coroutine and now dispatch back
// to our caller, based on TOS value
mp_int_t cause = MP_OBJ_SMALL_INT_VALUE(POP());
if (cause < 0) {
// A negative cause indicates unwind return
goto unwind_return;
} else {
// Otherwise it's an unwind jump and we must push as a raw
// number the number of exception handlers to unwind
PUSH((mp_obj_t)cause);
goto unwind_jump;
}
} else {
assert(mp_obj_is_exception_instance(TOP()));
RAISE(TOP());
}
DISPATCH();
ENTRY(MP_BC_GET_ITER):
MARK_EXC_IP_SELECTIVE();
SET_TOP(mp_getiter(TOP(), NULL));
DISPATCH();
// An iterator for a for-loop takes MP_OBJ_ITER_BUF_NSLOTS slots on
// the Python value stack. These slots are either used to store the
// iterator object itself, or the first slot is MP_OBJ_NULL and
// the second slot holds a reference to the iterator object.
ENTRY(MP_BC_GET_ITER_STACK): {
MARK_EXC_IP_SELECTIVE();
mp_obj_t obj = TOP();
mp_obj_iter_buf_t *iter_buf = (mp_obj_iter_buf_t*)sp;
sp += MP_OBJ_ITER_BUF_NSLOTS - 1;
obj = mp_getiter(obj, iter_buf);
if (obj != MP_OBJ_FROM_PTR(iter_buf)) {
// Iterator didn't use the stack so indicate that with MP_OBJ_NULL.
*(sp - MP_OBJ_ITER_BUF_NSLOTS + 1) = MP_OBJ_NULL;
*(sp - MP_OBJ_ITER_BUF_NSLOTS + 2) = obj;
}
DISPATCH();
}
ENTRY(MP_BC_FOR_ITER): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_ULABEL; // the jump offset if iteration finishes; for labels are always forward
code_state->sp = sp;
mp_obj_t obj;
if (*(sp - MP_OBJ_ITER_BUF_NSLOTS + 1) == MP_OBJ_NULL) {
obj = *(sp - MP_OBJ_ITER_BUF_NSLOTS + 2);
} else {
obj = MP_OBJ_FROM_PTR(&sp[-MP_OBJ_ITER_BUF_NSLOTS + 1]);
}
mp_obj_t value = mp_iternext_allow_raise(obj);
if (value == MP_OBJ_STOP_ITERATION) {
sp -= MP_OBJ_ITER_BUF_NSLOTS; // pop the exhausted iterator
ip += ulab; // jump to after for-block
} else {
PUSH(value); // push the next iteration value
#if MICROPY_PY_SYS_SETTRACE
// LINE event should trigger for every iteration so invalidate last trigger
if (code_state->frame) {
code_state->frame->lineno = 0;
}
#endif
}
DISPATCH();
}
ENTRY(MP_BC_POP_EXCEPT_JUMP): {
assert(exc_sp >= exc_stack);
POP_EXC_BLOCK();
DECODE_ULABEL;
ip += ulab;
DISPATCH_WITH_PEND_EXC_CHECK();
}
ENTRY(MP_BC_BUILD_TUPLE): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
sp -= unum - 1;
SET_TOP(mp_obj_new_tuple(unum, sp));
DISPATCH();
}
ENTRY(MP_BC_BUILD_LIST): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
sp -= unum - 1;
SET_TOP(mp_obj_new_list(unum, sp));
DISPATCH();
}
ENTRY(MP_BC_BUILD_MAP): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
PUSH(mp_obj_new_dict(unum));
DISPATCH();
}
ENTRY(MP_BC_STORE_MAP):
MARK_EXC_IP_SELECTIVE();
sp -= 2;
mp_obj_dict_store(sp[0], sp[2], sp[1]);
DISPATCH();
#if MICROPY_PY_BUILTINS_SET
ENTRY(MP_BC_BUILD_SET): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
sp -= unum - 1;
SET_TOP(mp_obj_new_set(unum, sp));
DISPATCH();
}
#endif
2013-10-16 20:57:49 +01:00
#if MICROPY_PY_BUILTINS_SLICE
ENTRY(MP_BC_BUILD_SLICE): {
MARK_EXC_IP_SELECTIVE();
mp_obj_t step = mp_const_none;
if (*ip++ == 3) {
// 3-argument slice includes step
step = POP();
}
mp_obj_t stop = POP();
mp_obj_t start = TOP();
SET_TOP(mp_obj_new_slice(start, stop, step));
DISPATCH();
}
#endif
ENTRY(MP_BC_STORE_COMP): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
mp_obj_t obj = sp[-(unum >> 2)];
if ((unum & 3) == 0) {
mp_obj_list_append(obj, sp[0]);
sp--;
} else if (!MICROPY_PY_BUILTINS_SET || (unum & 3) == 1) {
mp_obj_dict_store(obj, sp[0], sp[-1]);
sp -= 2;
#if MICROPY_PY_BUILTINS_SET
} else {
mp_obj_set_store(obj, sp[0]);
sp--;
#endif
}
DISPATCH();
}
ENTRY(MP_BC_UNPACK_SEQUENCE): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
mp_unpack_sequence(sp[0], unum, sp);
sp += unum - 1;
DISPATCH();
}
ENTRY(MP_BC_UNPACK_EX): {
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
mp_unpack_ex(sp[0], unum, sp);
sp += (unum & 0xff) + ((unum >> 8) & 0xff);
DISPATCH();
}
ENTRY(MP_BC_MAKE_FUNCTION): {
DECODE_PTR;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
PUSH(mp_make_function_from_raw_code(ptr, code_state->fun_bc->context, NULL));
DISPATCH();
}
ENTRY(MP_BC_MAKE_FUNCTION_DEFARGS): {
DECODE_PTR;
// Stack layout: def_tuple def_dict <- TOS
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
sp -= 1;
SET_TOP(mp_make_function_from_raw_code(ptr, code_state->fun_bc->context, sp));
DISPATCH();
}
ENTRY(MP_BC_MAKE_CLOSURE): {
DECODE_PTR;
size_t n_closed_over = *ip++;
// Stack layout: closed_overs <- TOS
sp -= 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 12:22:47 +01:00
SET_TOP(mp_make_closure_from_raw_code(ptr, code_state->fun_bc->context, n_closed_over, sp));
DISPATCH();
}
ENTRY(MP_BC_MAKE_CLOSURE_DEFARGS): {
DECODE_PTR;
size_t n_closed_over = *ip++;
// Stack layout: def_tuple def_dict closed_overs <- TOS
sp -= 2 + 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 12:22:47 +01:00
SET_TOP(mp_make_closure_from_raw_code(ptr, code_state->fun_bc->context, 0x100 | n_closed_over, sp));
DISPATCH();
}
ENTRY(MP_BC_CALL_FUNCTION): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
// unum & 0xff == n_positional
// (unum >> 8) & 0xff == n_keyword
sp -= (unum & 0xff) + ((unum >> 7) & 0x1fe);
#if MICROPY_STACKLESS
if (mp_obj_get_type(*sp) == &mp_type_fun_bc) {
code_state->ip = ip;
code_state->sp = sp;
code_state->exc_sp_idx = MP_CODE_STATE_EXC_SP_IDX_FROM_PTR(exc_stack, exc_sp);
mp_code_state_t *new_state = mp_obj_fun_bc_prepare_codestate(*sp, unum & 0xff, (unum >> 8) & 0xff, sp + 1);
#if !MICROPY_ENABLE_PYSTACK
if (new_state == NULL) {
// Couldn't allocate codestate on heap: in the strict case raise
// an exception, otherwise just fall through to stack allocation.
#if MICROPY_STACKLESS_STRICT
deep_recursion_error:
mp_raise_recursion_depth();
#endif
} else
#endif
{
new_state->prev = code_state;
code_state = new_state;
nlr_pop();
goto run_code_state;
}
}
#endif
SET_TOP(mp_call_function_n_kw(*sp, unum & 0xff, (unum >> 8) & 0xff, sp + 1));
DISPATCH();
}
ENTRY(MP_BC_CALL_FUNCTION_VAR_KW): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
// unum & 0xff == n_positional
// (unum >> 8) & 0xff == n_keyword
2017-05-29 08:08:14 +01:00
// We have following stack layout here:
// fun arg0 arg1 ... kw0 val0 kw1 val1 ... seq dict <- TOS
sp -= (unum & 0xff) + ((unum >> 7) & 0x1fe) + 2;
#if MICROPY_STACKLESS
if (mp_obj_get_type(*sp) == &mp_type_fun_bc) {
code_state->ip = ip;
code_state->sp = sp;
code_state->exc_sp_idx = MP_CODE_STATE_EXC_SP_IDX_FROM_PTR(exc_stack, exc_sp);
mp_call_args_t out_args;
mp_call_prepare_args_n_kw_var(false, unum, sp, &out_args);
mp_code_state_t *new_state = mp_obj_fun_bc_prepare_codestate(out_args.fun,
out_args.n_args, out_args.n_kw, out_args.args);
#if !MICROPY_ENABLE_PYSTACK
// Freeing args at this point does not follow a LIFO order so only do it if
// pystack is not enabled. For pystack, they are freed when code_state is.
mp_nonlocal_free(out_args.args, out_args.n_alloc * sizeof(mp_obj_t));
#endif
#if !MICROPY_ENABLE_PYSTACK
if (new_state == NULL) {
// Couldn't allocate codestate on heap: in the strict case raise
// an exception, otherwise just fall through to stack allocation.
#if MICROPY_STACKLESS_STRICT
goto deep_recursion_error;
#endif
} else
#endif
{
new_state->prev = code_state;
code_state = new_state;
nlr_pop();
goto run_code_state;
}
}
#endif
SET_TOP(mp_call_method_n_kw_var(false, unum, sp));
DISPATCH();
}
ENTRY(MP_BC_CALL_METHOD): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
// unum & 0xff == n_positional
// (unum >> 8) & 0xff == n_keyword
sp -= (unum & 0xff) + ((unum >> 7) & 0x1fe) + 1;
#if MICROPY_STACKLESS
if (mp_obj_get_type(*sp) == &mp_type_fun_bc) {
code_state->ip = ip;
code_state->sp = sp;
code_state->exc_sp_idx = MP_CODE_STATE_EXC_SP_IDX_FROM_PTR(exc_stack, exc_sp);
size_t n_args = unum & 0xff;
size_t n_kw = (unum >> 8) & 0xff;
int adjust = (sp[1] == MP_OBJ_NULL) ? 0 : 1;
mp_code_state_t *new_state = mp_obj_fun_bc_prepare_codestate(*sp, n_args + adjust, n_kw, sp + 2 - adjust);
#if !MICROPY_ENABLE_PYSTACK
if (new_state == NULL) {
// Couldn't allocate codestate on heap: in the strict case raise
// an exception, otherwise just fall through to stack allocation.
#if MICROPY_STACKLESS_STRICT
goto deep_recursion_error;
#endif
} else
#endif
{
new_state->prev = code_state;
code_state = new_state;
nlr_pop();
goto run_code_state;
}
}
#endif
SET_TOP(mp_call_method_n_kw(unum & 0xff, (unum >> 8) & 0xff, sp));
DISPATCH();
}
ENTRY(MP_BC_CALL_METHOD_VAR_KW): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_UINT;
// unum & 0xff == n_positional
// (unum >> 8) & 0xff == n_keyword
2017-05-29 08:08:14 +01:00
// We have following stack layout here:
// fun self arg0 arg1 ... kw0 val0 kw1 val1 ... seq dict <- TOS
sp -= (unum & 0xff) + ((unum >> 7) & 0x1fe) + 3;
#if MICROPY_STACKLESS
if (mp_obj_get_type(*sp) == &mp_type_fun_bc) {
code_state->ip = ip;
code_state->sp = sp;
code_state->exc_sp_idx = MP_CODE_STATE_EXC_SP_IDX_FROM_PTR(exc_stack, exc_sp);
mp_call_args_t out_args;
mp_call_prepare_args_n_kw_var(true, unum, sp, &out_args);
mp_code_state_t *new_state = mp_obj_fun_bc_prepare_codestate(out_args.fun,
out_args.n_args, out_args.n_kw, out_args.args);
#if !MICROPY_ENABLE_PYSTACK
// Freeing args at this point does not follow a LIFO order so only do it if
// pystack is not enabled. For pystack, they are freed when code_state is.
mp_nonlocal_free(out_args.args, out_args.n_alloc * sizeof(mp_obj_t));
#endif
#if !MICROPY_ENABLE_PYSTACK
if (new_state == NULL) {
// Couldn't allocate codestate on heap: in the strict case raise
// an exception, otherwise just fall through to stack allocation.
#if MICROPY_STACKLESS_STRICT
goto deep_recursion_error;
#endif
} else
#endif
{
new_state->prev = code_state;
code_state = new_state;
nlr_pop();
goto run_code_state;
}
}
#endif
SET_TOP(mp_call_method_n_kw_var(true, unum, sp));
DISPATCH();
}
ENTRY(MP_BC_RETURN_VALUE):
MARK_EXC_IP_SELECTIVE();
unwind_return:
// Search for and execute finally handlers that aren't already active
while (exc_sp >= exc_stack) {
if (MP_TAGPTR_TAG1(exc_sp->val_sp)) {
if (exc_sp->handler > ip) {
// Found a finally handler that isn't active; run it.
// Getting here the stack looks like:
// (..., X, [iter0, iter1, ...,] ret_val)
// where X is pointed to by exc_sp->val_sp and in the case
// of a "with" block contains the context manager info.
// There may be 0 or more for-iterators between X and the
// return value, and these must be removed before control can
// pass to the finally code. We simply copy the ret_value down
// over these iterators, if they exist. If they don't then the
// following is a null operation.
mp_obj_t *finally_sp = MP_TAGPTR_PTR(exc_sp->val_sp);
finally_sp[1] = sp[0];
sp = &finally_sp[1];
// We're going to run "finally" code as a coroutine
// (not calling it recursively). Set up a sentinel
// on a stack so it can return back to us when it is
// done (when WITH_CLEANUP or END_FINALLY reached).
PUSH(MP_OBJ_NEW_SMALL_INT(-1));
ip = exc_sp->handler;
goto dispatch_loop;
} else {
// Found a finally handler that is already active; cancel it.
CANCEL_ACTIVE_FINALLY(sp);
}
}
POP_EXC_BLOCK();
}
nlr_pop();
code_state->sp = sp;
assert(exc_sp == exc_stack - 1);
MICROPY_VM_HOOK_RETURN
#if MICROPY_STACKLESS
if (code_state->prev != NULL) {
mp_obj_t res = *sp;
mp_globals_set(code_state->old_globals);
mp_code_state_t *new_code_state = code_state->prev;
#if MICROPY_ENABLE_PYSTACK
// Free code_state, and args allocated by mp_call_prepare_args_n_kw_var
// (The latter is implicitly freed when using pystack due to its LIFO nature.)
// The sizeof in the following statement does not include the size of the variable
// part of the struct. This arg is anyway not used if pystack is enabled.
mp_nonlocal_free(code_state, sizeof(mp_code_state_t));
#endif
code_state = new_code_state;
*code_state->sp = res;
goto run_code_state_from_return;
}
#endif
FRAME_LEAVE();
return MP_VM_RETURN_NORMAL;
ENTRY(MP_BC_RAISE_LAST): {
MARK_EXC_IP_SELECTIVE();
// search for the inner-most previous exception, to reraise it
mp_obj_t obj = MP_OBJ_NULL;
for (mp_exc_stack_t *e = exc_sp; e >= exc_stack; --e) {
if (e->prev_exc != NULL) {
obj = MP_OBJ_FROM_PTR(e->prev_exc);
break;
}
}
if (obj == MP_OBJ_NULL) {
obj = mp_obj_new_exception_msg(&mp_type_RuntimeError, MP_ERROR_TEXT("no active exception to reraise"));
}
RAISE(obj);
}
ENTRY(MP_BC_RAISE_OBJ): {
MARK_EXC_IP_SELECTIVE();
mp_obj_t obj = mp_make_raise_obj(TOP());
RAISE(obj);
}
ENTRY(MP_BC_RAISE_FROM): {
MARK_EXC_IP_SELECTIVE();
mp_warning(NULL, "exception chaining not supported");
sp--; // ignore (pop) "from" argument
mp_obj_t obj = mp_make_raise_obj(TOP());
RAISE(obj);
}
ENTRY(MP_BC_YIELD_VALUE):
2014-03-26 15:36:12 +00:00
yield:
nlr_pop();
code_state->ip = ip;
code_state->sp = sp;
code_state->exc_sp_idx = MP_CODE_STATE_EXC_SP_IDX_FROM_PTR(exc_stack, exc_sp);
FRAME_LEAVE();
return MP_VM_RETURN_YIELD;
ENTRY(MP_BC_YIELD_FROM): {
MARK_EXC_IP_SELECTIVE();
//#define EXC_MATCH(exc, type) mp_obj_is_type(exc, type)
2014-03-26 15:36:12 +00:00
#define EXC_MATCH(exc, type) mp_obj_exception_match(exc, type)
#define GENERATOR_EXIT_IF_NEEDED(t) if (t != MP_OBJ_NULL && EXC_MATCH(t, MP_OBJ_FROM_PTR(&mp_type_GeneratorExit))) { mp_obj_t raise_t = mp_make_raise_obj(t); RAISE(raise_t); }
mp_vm_return_kind_t ret_kind;
mp_obj_t send_value = POP();
mp_obj_t t_exc = MP_OBJ_NULL;
mp_obj_t ret_value;
code_state->sp = sp; // Save sp because it's needed if mp_resume raises StopIteration
if (inject_exc != MP_OBJ_NULL) {
t_exc = inject_exc;
inject_exc = MP_OBJ_NULL;
ret_kind = mp_resume(TOP(), MP_OBJ_NULL, t_exc, &ret_value);
} else {
ret_kind = mp_resume(TOP(), send_value, MP_OBJ_NULL, &ret_value);
}
2014-03-26 15:36:12 +00:00
if (ret_kind == MP_VM_RETURN_YIELD) {
ip--;
PUSH(ret_value);
goto yield;
} else if (ret_kind == MP_VM_RETURN_NORMAL) {
// The generator has finished, and returned a value via StopIteration
// Replace exhausted generator with the returned value
SET_TOP(ret_value);
// If we injected GeneratorExit downstream, then even
// if it was swallowed, we re-raise GeneratorExit
GENERATOR_EXIT_IF_NEEDED(t_exc);
DISPATCH();
} else {
assert(ret_kind == MP_VM_RETURN_EXCEPTION);
assert(!EXC_MATCH(ret_value, MP_OBJ_FROM_PTR(&mp_type_StopIteration)));
// Pop exhausted gen
sp--;
RAISE(ret_value);
2014-03-26 15:36:12 +00:00
}
}
2014-03-26 15:36:12 +00:00
ENTRY(MP_BC_IMPORT_NAME): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_obj_t obj = POP();
SET_TOP(mp_import_name(qst, obj, TOP()));
DISPATCH();
}
ENTRY(MP_BC_IMPORT_FROM): {
FRAME_UPDATE();
MARK_EXC_IP_SELECTIVE();
DECODE_QSTR;
mp_obj_t obj = mp_import_from(TOP(), qst);
PUSH(obj);
DISPATCH();
}
ENTRY(MP_BC_IMPORT_STAR):
MARK_EXC_IP_SELECTIVE();
mp_import_all(POP());
DISPATCH();
#if MICROPY_OPT_COMPUTED_GOTO
ENTRY(MP_BC_LOAD_CONST_SMALL_INT_MULTI):
PUSH(MP_OBJ_NEW_SMALL_INT((mp_int_t)ip[-1] - MP_BC_LOAD_CONST_SMALL_INT_MULTI - MP_BC_LOAD_CONST_SMALL_INT_MULTI_EXCESS));
DISPATCH();
ENTRY(MP_BC_LOAD_FAST_MULTI):
obj_shared = fastn[MP_BC_LOAD_FAST_MULTI - (mp_int_t)ip[-1]];
goto load_check;
ENTRY(MP_BC_STORE_FAST_MULTI):
fastn[MP_BC_STORE_FAST_MULTI - (mp_int_t)ip[-1]] = POP();
DISPATCH();
ENTRY(MP_BC_UNARY_OP_MULTI):
MARK_EXC_IP_SELECTIVE();
SET_TOP(mp_unary_op(ip[-1] - MP_BC_UNARY_OP_MULTI, TOP()));
DISPATCH();
ENTRY(MP_BC_BINARY_OP_MULTI): {
MARK_EXC_IP_SELECTIVE();
mp_obj_t rhs = POP();
mp_obj_t lhs = TOP();
SET_TOP(mp_binary_op(ip[-1] - MP_BC_BINARY_OP_MULTI, lhs, rhs));
DISPATCH();
}
ENTRY_DEFAULT:
MARK_EXC_IP_SELECTIVE();
#else
ENTRY_DEFAULT:
if (ip[-1] < MP_BC_LOAD_CONST_SMALL_INT_MULTI + MP_BC_LOAD_CONST_SMALL_INT_MULTI_NUM) {
PUSH(MP_OBJ_NEW_SMALL_INT((mp_int_t)ip[-1] - MP_BC_LOAD_CONST_SMALL_INT_MULTI - MP_BC_LOAD_CONST_SMALL_INT_MULTI_EXCESS));
DISPATCH();
} else if (ip[-1] < MP_BC_LOAD_FAST_MULTI + MP_BC_LOAD_FAST_MULTI_NUM) {
obj_shared = fastn[MP_BC_LOAD_FAST_MULTI - (mp_int_t)ip[-1]];
goto load_check;
} else if (ip[-1] < MP_BC_STORE_FAST_MULTI + MP_BC_STORE_FAST_MULTI_NUM) {
fastn[MP_BC_STORE_FAST_MULTI - (mp_int_t)ip[-1]] = POP();
DISPATCH();
} else if (ip[-1] < MP_BC_UNARY_OP_MULTI + MP_BC_UNARY_OP_MULTI_NUM) {
SET_TOP(mp_unary_op(ip[-1] - MP_BC_UNARY_OP_MULTI, TOP()));
DISPATCH();
} else if (ip[-1] < MP_BC_BINARY_OP_MULTI + MP_BC_BINARY_OP_MULTI_NUM) {
mp_obj_t rhs = POP();
mp_obj_t lhs = TOP();
SET_TOP(mp_binary_op(ip[-1] - MP_BC_BINARY_OP_MULTI, lhs, rhs));
DISPATCH();
} else
#endif
{
mp_obj_t obj = mp_obj_new_exception_msg(&mp_type_NotImplementedError, MP_ERROR_TEXT("opcode"));
nlr_pop();
code_state->state[0] = obj;
FRAME_LEAVE();
return MP_VM_RETURN_EXCEPTION;
}
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
#if !MICROPY_OPT_COMPUTED_GOTO
} // switch
#endif
pending_exception_check:
MICROPY_VM_HOOK_LOOP
#if MICROPY_ENABLE_SCHEDULER
// This is an inlined variant of mp_handle_pending
if (MP_STATE_VM(sched_state) == MP_SCHED_PENDING) {
mp_uint_t atomic_state = MICROPY_BEGIN_ATOMIC_SECTION();
// Re-check state is still pending now that we're in the atomic section.
if (MP_STATE_VM(sched_state) == MP_SCHED_PENDING) {
MARK_EXC_IP_SELECTIVE();
mp_obj_t obj = MP_STATE_THREAD(mp_pending_exception);
if (obj != MP_OBJ_NULL) {
MP_STATE_THREAD(mp_pending_exception) = MP_OBJ_NULL;
if (!mp_sched_num_pending()) {
MP_STATE_VM(sched_state) = MP_SCHED_IDLE;
}
MICROPY_END_ATOMIC_SECTION(atomic_state);
RAISE(obj);
}
mp_handle_pending_tail(atomic_state);
} else {
MICROPY_END_ATOMIC_SECTION(atomic_state);
}
}
#else
// This is an inlined variant of mp_handle_pending
if (MP_STATE_THREAD(mp_pending_exception) != MP_OBJ_NULL) {
MARK_EXC_IP_SELECTIVE();
mp_obj_t obj = MP_STATE_THREAD(mp_pending_exception);
MP_STATE_THREAD(mp_pending_exception) = MP_OBJ_NULL;
RAISE(obj);
}
#endif
#if MICROPY_PY_THREAD_GIL
#if MICROPY_PY_THREAD_GIL_VM_DIVISOR
if (--gil_divisor == 0)
#endif
{
#if MICROPY_PY_THREAD_GIL_VM_DIVISOR
gil_divisor = MICROPY_PY_THREAD_GIL_VM_DIVISOR;
#endif
#if MICROPY_ENABLE_SCHEDULER
// can only switch threads if the scheduler is unlocked
if (MP_STATE_VM(sched_state) == MP_SCHED_IDLE)
#endif
{
MP_THREAD_GIL_EXIT();
MP_THREAD_GIL_ENTER();
}
}
#endif
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
} // for loop
} else {
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
exception_handler:
// exception occurred
#if MICROPY_PY_SYS_EXC_INFO
MP_STATE_VM(cur_exception) = nlr.ret_val;
#endif
#if SELECTIVE_EXC_IP
// with selective ip, we store the ip 1 byte past the opcode, so move ptr back
code_state->ip -= 1;
#endif
if (mp_obj_is_subclass_fast(MP_OBJ_FROM_PTR(((mp_obj_base_t*)nlr.ret_val)->type), MP_OBJ_FROM_PTR(&mp_type_StopIteration))) {
if (code_state->ip) {
// check if it's a StopIteration within a for block
if (*code_state->ip == MP_BC_FOR_ITER) {
const byte *ip = code_state->ip + 1;
DECODE_ULABEL; // the jump offset if iteration finishes; for labels are always forward
code_state->ip = ip + ulab; // jump to after for-block
code_state->sp -= MP_OBJ_ITER_BUF_NSLOTS; // pop the exhausted iterator
goto outer_dispatch_loop; // continue with dispatch loop
} else if (*code_state->ip == MP_BC_YIELD_FROM) {
// StopIteration inside yield from call means return a value of
// yield from, so inject exception's value as yield from's result
// (Instead of stack pop then push we just replace exhausted gen with value)
*code_state->sp = mp_obj_exception_get_value(MP_OBJ_FROM_PTR(nlr.ret_val));
code_state->ip++; // yield from is over, move to next instruction
goto outer_dispatch_loop; // continue with dispatch loop
}
}
}
#if MICROPY_PY_SYS_SETTRACE
// Exceptions are traced here
if (mp_obj_is_subclass_fast(MP_OBJ_FROM_PTR(((mp_obj_base_t*)nlr.ret_val)->type), MP_OBJ_FROM_PTR(&mp_type_Exception))) {
TRACE_TICK(code_state->ip, code_state->sp, true /* yes, it's an exception */);
}
#endif
#if MICROPY_STACKLESS
unwind_loop:
#endif
// Set traceback info (file and line number) where the exception occurred, but not for:
// - constant GeneratorExit object, because it's const
// - exceptions re-raised by END_FINALLY
// - exceptions re-raised explicitly by "raise"
if (nlr.ret_val != &mp_const_GeneratorExit_obj
&& *code_state->ip != MP_BC_END_FINALLY
&& *code_state->ip != MP_BC_RAISE_LAST) {
const byte *ip = code_state->fun_bc->bytecode;
MP_BC_PRELUDE_SIG_DECODE(ip);
MP_BC_PRELUDE_SIZE_DECODE(ip);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
const byte *line_info_top = ip + n_info;
const byte *bytecode_start = ip + n_info + n_cell;
size_t bc = code_state->ip - bytecode_start;
qstr block_name = mp_decode_uint_value(ip);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
for (size_t i = 0; i < 1 + n_pos_args + n_kwonly_args; ++i) {
ip = mp_decode_uint_skip(ip);
}
#if MICROPY_EMIT_BYTECODE_USES_QSTR_TABLE
block_name = code_state->fun_bc->context->constants.qstr_table[block_name];
qstr source_file = code_state->fun_bc->context->constants.qstr_table[0];
#else
qstr source_file = code_state->fun_bc->context->constants.source_file;
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t source_line = mp_bytecode_get_source_line(ip, line_info_top, bc);
mp_obj_exception_add_traceback(MP_OBJ_FROM_PTR(nlr.ret_val), source_file, source_line, block_name);
}
while (exc_sp >= exc_stack && exc_sp->handler <= code_state->ip) {
// nested exception
assert(exc_sp >= exc_stack);
// TODO make a proper message for nested exception
// at the moment we are just raising the very last exception (the one that caused the nested exception)
// move up to previous exception handler
POP_EXC_BLOCK();
}
if (exc_sp >= exc_stack) {
// catch exception and pass to byte code
code_state->ip = exc_sp->handler;
py: Tidy up variables in VM, probably fixes subtle bugs. Things get tricky when using the nlr code to catch exceptions. Need to ensure that the variables (stack layout) in the exception handler are the same as in the bit protected by the exception handler. Prior to this patch there were a few bugs. 1) The constant mp_const_MemoryError_obj was being preloaded to a specific location on the stack at the start of the function. But this location on the stack was being overwritten in the opcode loop (since it didn't think that variable would ever be referenced again), and so when an exception occurred, the variable holding the address of MemoryError was corrupt. 2) The FOR_ITER opcode detection in the exception handler used sp, which may or may not contain the right value coming out of the main opcode loop. With this patch there is a clear separation of variables used in the opcode loop and in the exception handler (should fix issue (2) above). Furthermore, nlr_raise is no longer used in the opcode loop. Instead, it jumps directly into the exception handler. This tells the C compiler more about the possible code flow, and means that it should have the same stack layout for the exception handler. This should fix issue (1) above. Indeed, the generated (ARM) assembler has been checked explicitly, and with 'goto exception_handler', the problem with &MemoryError is fixed. This may now fix problems with rge-sm, and probably many other subtle bugs yet to show themselves. Incidentally, rge-sm now passes on pyboard (with a reduced range of integration)! Main lesson: nlr is tricky. Don't use nlr_push unless you know what you are doing! Luckily, it's not used in many places. Using nlr_raise/jump is fine.
2014-04-17 16:50:23 +01:00
mp_obj_t *sp = MP_TAGPTR_PTR(exc_sp->val_sp);
2014-03-30 00:54:48 +00:00
// save this exception in the stack so it can be used in a reraise, if needed
exc_sp->prev_exc = nlr.ret_val;
// push exception object so it can be handled by bytecode
PUSH(MP_OBJ_FROM_PTR(nlr.ret_val));
code_state->sp = sp;
#if MICROPY_STACKLESS
} else if (code_state->prev != NULL) {
mp_globals_set(code_state->old_globals);
mp_code_state_t *new_code_state = code_state->prev;
#if MICROPY_ENABLE_PYSTACK
// Free code_state, and args allocated by mp_call_prepare_args_n_kw_var
// (The latter is implicitly freed when using pystack due to its LIFO nature.)
// The sizeof in the following statement does not include the size of the variable
// part of the struct. This arg is anyway not used if pystack is enabled.
mp_nonlocal_free(code_state, sizeof(mp_code_state_t));
#endif
code_state = new_code_state;
size_t n_state = code_state->n_state;
fastn = &code_state->state[n_state - 1];
exc_stack = (mp_exc_stack_t*)(code_state->state + n_state);
// variables that are visible to the exception handler (declared volatile)
exc_sp = MP_CODE_STATE_EXC_SP_IDX_TO_PTR(exc_stack, code_state->exc_sp_idx); // stack grows up, exc_sp points to top of stack
goto unwind_loop;
#endif
} else {
// propagate exception to higher level
// Note: ip and sp don't have usable values at this point
code_state->state[0] = MP_OBJ_FROM_PTR(nlr.ret_val); // put exception here because sp is invalid
FRAME_LEAVE();
return MP_VM_RETURN_EXCEPTION;
}
2013-10-04 19:53:11 +01:00
}
}
}