2019-08-28 08:03:08 +01:00
|
|
|
#!/usr/bin/env python3
|
|
|
|
#
|
|
|
|
# This file is part of the MicroPython project, http://micropython.org/
|
|
|
|
#
|
|
|
|
# The MIT License (MIT)
|
|
|
|
#
|
|
|
|
# Copyright (c) 2019 Damien P. George
|
|
|
|
#
|
|
|
|
# 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.
|
|
|
|
|
|
|
|
"""
|
|
|
|
Link .o files to .mpy
|
|
|
|
"""
|
|
|
|
|
|
|
|
import sys, os, struct, re
|
|
|
|
from elftools.elf import elffile
|
|
|
|
|
|
|
|
sys.path.append(os.path.dirname(__file__) + "/../py")
|
|
|
|
import makeqstrdata as qstrutil
|
|
|
|
|
|
|
|
# MicroPython constants
|
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
|
|
|
MPY_VERSION = 6
|
2022-09-17 14:57:12 +01:00
|
|
|
MPY_SUB_VERSION = 1
|
2022-05-26 01:51:29 +01:00
|
|
|
MP_CODE_BYTECODE = 2
|
|
|
|
MP_CODE_NATIVE_VIPER = 4
|
2019-08-28 08:03:08 +01:00
|
|
|
MP_NATIVE_ARCH_X86 = 1
|
|
|
|
MP_NATIVE_ARCH_X64 = 2
|
2022-05-23 13:50:34 +01:00
|
|
|
MP_NATIVE_ARCH_ARMV6M = 4
|
2019-08-28 08:03:08 +01:00
|
|
|
MP_NATIVE_ARCH_ARMV7M = 5
|
|
|
|
MP_NATIVE_ARCH_ARMV7EMSP = 7
|
|
|
|
MP_NATIVE_ARCH_ARMV7EMDP = 8
|
|
|
|
MP_NATIVE_ARCH_XTENSA = 9
|
|
|
|
MP_NATIVE_ARCH_XTENSAWIN = 10
|
2022-05-26 01:51:29 +01:00
|
|
|
MP_PERSISTENT_OBJ_STR = 5
|
2019-08-28 08:03:08 +01:00
|
|
|
MP_SCOPE_FLAG_VIPERRELOC = 0x10
|
|
|
|
MP_SCOPE_FLAG_VIPERRODATA = 0x20
|
|
|
|
MP_SCOPE_FLAG_VIPERBSS = 0x40
|
|
|
|
MP_SMALL_INT_BITS = 31
|
|
|
|
|
|
|
|
# ELF constants
|
|
|
|
R_386_32 = 1
|
|
|
|
R_X86_64_64 = 1
|
|
|
|
R_XTENSA_32 = 1
|
|
|
|
R_386_PC32 = 2
|
|
|
|
R_X86_64_PC32 = 2
|
|
|
|
R_ARM_ABS32 = 2
|
|
|
|
R_386_GOT32 = 3
|
|
|
|
R_ARM_REL32 = 3
|
|
|
|
R_386_PLT32 = 4
|
|
|
|
R_X86_64_PLT32 = 4
|
|
|
|
R_XTENSA_PLT = 6
|
|
|
|
R_386_GOTOFF = 9
|
|
|
|
R_386_GOTPC = 10
|
|
|
|
R_ARM_THM_CALL = 10
|
|
|
|
R_XTENSA_DIFF32 = 19
|
|
|
|
R_XTENSA_SLOT0_OP = 20
|
|
|
|
R_ARM_BASE_PREL = 25 # aka R_ARM_GOTPC
|
|
|
|
R_ARM_GOT_BREL = 26 # aka R_ARM_GOT32
|
|
|
|
R_ARM_THM_JUMP24 = 30
|
2020-11-14 07:54:26 +00:00
|
|
|
R_X86_64_GOTPCREL = 9
|
2019-08-28 08:03:08 +01:00
|
|
|
R_X86_64_REX_GOTPCRELX = 42
|
|
|
|
R_386_GOT32X = 43
|
2022-06-09 02:57:22 +01:00
|
|
|
R_XTENSA_PDIFF32 = 59
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
################################################################################
|
|
|
|
# Architecture configuration
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def asm_jump_x86(entry):
|
|
|
|
return struct.pack("<BI", 0xE9, entry - 5)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
2022-05-23 13:50:34 +01:00
|
|
|
def asm_jump_thumb(entry):
|
|
|
|
# Only signed values that fit in 12 bits are supported
|
|
|
|
b_off = entry - 4
|
|
|
|
assert b_off >> 11 == 0 or b_off >> 11 == -1, b_off
|
|
|
|
return struct.pack("<H", 0xE000 | (b_off >> 1 & 0x07FF))
|
|
|
|
|
|
|
|
|
|
|
|
def asm_jump_thumb2(entry):
|
2019-08-28 08:03:08 +01:00
|
|
|
b_off = entry - 4
|
|
|
|
if b_off >> 11 == 0 or b_off >> 11 == -1:
|
|
|
|
# Signed value fits in 12 bits
|
|
|
|
b0 = 0xE000 | (b_off >> 1 & 0x07FF)
|
|
|
|
b1 = 0
|
|
|
|
else:
|
|
|
|
# Use large jump
|
|
|
|
b0 = 0xF000 | (b_off >> 12 & 0x07FF)
|
|
|
|
b1 = 0xB800 | (b_off >> 1 & 0x7FF)
|
|
|
|
return struct.pack("<HH", b0, b1)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def asm_jump_xtensa(entry):
|
|
|
|
jump_offset = entry - 4
|
|
|
|
jump_op = jump_offset << 6 | 6
|
|
|
|
return struct.pack("<BH", jump_op & 0xFF, jump_op >> 8)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
class ArchData:
|
2022-05-26 01:51:29 +01:00
|
|
|
def __init__(self, name, mpy_feature, word_size, arch_got, asm_jump, *, separate_rodata=False):
|
2019-08-28 08:03:08 +01:00
|
|
|
self.name = name
|
|
|
|
self.mpy_feature = mpy_feature
|
2022-05-26 01:51:29 +01:00
|
|
|
self.qstr_entry_size = 2
|
2019-08-28 08:03:08 +01:00
|
|
|
self.word_size = word_size
|
|
|
|
self.arch_got = arch_got
|
|
|
|
self.asm_jump = asm_jump
|
2022-05-26 01:51:29 +01:00
|
|
|
self.separate_rodata = separate_rodata
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
ARCH_DATA = {
|
|
|
|
"x86": ArchData(
|
|
|
|
"EM_386",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_X86 << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_386_PC32, R_386_GOT32, R_386_GOT32X),
|
|
|
|
asm_jump_x86,
|
|
|
|
),
|
|
|
|
"x64": ArchData(
|
|
|
|
"EM_X86_64",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_X64 << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
8,
|
2020-11-14 07:54:26 +00:00
|
|
|
(R_X86_64_GOTPCREL, R_X86_64_REX_GOTPCRELX),
|
2019-08-28 08:03:08 +01:00
|
|
|
asm_jump_x86,
|
|
|
|
),
|
2022-05-23 13:50:34 +01:00
|
|
|
"armv6m": ArchData(
|
|
|
|
"EM_ARM",
|
|
|
|
MP_NATIVE_ARCH_ARMV6M << 2,
|
|
|
|
4,
|
|
|
|
(R_ARM_GOT_BREL,),
|
|
|
|
asm_jump_thumb,
|
|
|
|
),
|
2019-08-28 08:03:08 +01:00
|
|
|
"armv7m": ArchData(
|
|
|
|
"EM_ARM",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_ARMV7M << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_ARM_GOT_BREL,),
|
2022-05-23 13:50:34 +01:00
|
|
|
asm_jump_thumb2,
|
2019-08-28 08:03:08 +01:00
|
|
|
),
|
|
|
|
"armv7emsp": ArchData(
|
|
|
|
"EM_ARM",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_ARMV7EMSP << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_ARM_GOT_BREL,),
|
2022-05-23 13:50:34 +01:00
|
|
|
asm_jump_thumb2,
|
2019-08-28 08:03:08 +01:00
|
|
|
),
|
|
|
|
"armv7emdp": ArchData(
|
|
|
|
"EM_ARM",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_ARMV7EMDP << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_ARM_GOT_BREL,),
|
2022-05-23 13:50:34 +01:00
|
|
|
asm_jump_thumb2,
|
2019-08-28 08:03:08 +01:00
|
|
|
),
|
|
|
|
"xtensa": ArchData(
|
|
|
|
"EM_XTENSA",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_XTENSA << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_XTENSA_32, R_XTENSA_PLT),
|
|
|
|
asm_jump_xtensa,
|
|
|
|
),
|
|
|
|
"xtensawin": ArchData(
|
|
|
|
"EM_XTENSA",
|
2022-05-16 10:20:52 +01:00
|
|
|
MP_NATIVE_ARCH_XTENSAWIN << 2,
|
2019-08-28 08:03:08 +01:00
|
|
|
4,
|
|
|
|
(R_XTENSA_32, R_XTENSA_PLT),
|
|
|
|
asm_jump_xtensa,
|
2022-05-26 01:51:29 +01:00
|
|
|
separate_rodata=True,
|
2019-08-28 08:03:08 +01:00
|
|
|
),
|
|
|
|
}
|
|
|
|
|
|
|
|
################################################################################
|
|
|
|
# Helper functions
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def align_to(value, align):
|
|
|
|
return (value + align - 1) & ~(align - 1)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def unpack_u24le(data, offset):
|
|
|
|
return data[offset] | data[offset + 1] << 8 | data[offset + 2] << 16
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def pack_u24le(data, offset, value):
|
|
|
|
data[offset] = value & 0xFF
|
|
|
|
data[offset + 1] = value >> 8 & 0xFF
|
|
|
|
data[offset + 2] = value >> 16 & 0xFF
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def xxd(text):
|
|
|
|
for i in range(0, len(text), 16):
|
|
|
|
print("{:08x}:".format(i), end="")
|
|
|
|
for j in range(4):
|
|
|
|
off = i + j * 4
|
|
|
|
if off < len(text):
|
|
|
|
d = int.from_bytes(text[off : off + 4], "little")
|
|
|
|
print(" {:08x}".format(d), end="")
|
|
|
|
print()
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
# Smaller numbers are enabled first
|
|
|
|
LOG_LEVEL_1 = 1
|
|
|
|
LOG_LEVEL_2 = 2
|
|
|
|
LOG_LEVEL_3 = 3
|
|
|
|
log_level = LOG_LEVEL_1
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def log(level, msg):
|
|
|
|
if level <= log_level:
|
|
|
|
print(msg)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
################################################################################
|
|
|
|
# Qstr extraction
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def extract_qstrs(source_files):
|
|
|
|
def read_qstrs(f):
|
|
|
|
with open(f) as f:
|
|
|
|
vals = set()
|
|
|
|
objs = set()
|
|
|
|
for line in f:
|
|
|
|
while line:
|
|
|
|
m = re.search(r"MP_OBJ_NEW_QSTR\((MP_QSTR_[A-Za-z0-9_]*)\)", line)
|
|
|
|
if m:
|
|
|
|
objs.add(m.group(1))
|
|
|
|
else:
|
|
|
|
m = re.search(r"MP_QSTR_[A-Za-z0-9_]*", line)
|
|
|
|
if m:
|
|
|
|
vals.add(m.group())
|
|
|
|
if m:
|
|
|
|
s = m.span()
|
|
|
|
line = line[: s[0]] + line[s[1] :]
|
|
|
|
else:
|
|
|
|
line = ""
|
|
|
|
return vals, objs
|
|
|
|
|
|
|
|
static_qstrs = ["MP_QSTR_" + qstrutil.qstr_escape(q) for q in qstrutil.static_qstr_list]
|
|
|
|
|
|
|
|
qstr_vals = set()
|
|
|
|
qstr_objs = set()
|
|
|
|
for f in source_files:
|
|
|
|
vals, objs = read_qstrs(f)
|
|
|
|
qstr_vals.update(vals)
|
|
|
|
qstr_objs.update(objs)
|
|
|
|
qstr_vals.difference_update(static_qstrs)
|
|
|
|
|
|
|
|
return static_qstrs, qstr_vals, qstr_objs
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
################################################################################
|
|
|
|
# Linker
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
class LinkError(Exception):
|
|
|
|
pass
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
class Section:
|
|
|
|
def __init__(self, name, data, alignment, filename=None):
|
|
|
|
self.filename = filename
|
|
|
|
self.name = name
|
|
|
|
self.data = data
|
|
|
|
self.alignment = alignment
|
|
|
|
self.addr = 0
|
|
|
|
self.reloc = []
|
|
|
|
|
|
|
|
@staticmethod
|
|
|
|
def from_elfsec(elfsec, filename):
|
|
|
|
assert elfsec.header.sh_addr == 0
|
|
|
|
return Section(elfsec.name, elfsec.data(), elfsec.data_alignment, filename)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
class GOTEntry:
|
|
|
|
def __init__(self, name, sym, link_addr=0):
|
|
|
|
self.name = name
|
|
|
|
self.sym = sym
|
|
|
|
self.offset = None
|
|
|
|
self.link_addr = link_addr
|
|
|
|
|
|
|
|
def isexternal(self):
|
|
|
|
return self.sec_name.startswith(".external")
|
|
|
|
|
|
|
|
def istext(self):
|
|
|
|
return self.sec_name.startswith(".text")
|
|
|
|
|
|
|
|
def isrodata(self):
|
|
|
|
return self.sec_name.startswith((".rodata", ".data.rel.ro"))
|
|
|
|
|
|
|
|
def isbss(self):
|
|
|
|
return self.sec_name.startswith(".bss")
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
class LiteralEntry:
|
|
|
|
def __init__(self, value, offset):
|
|
|
|
self.value = value
|
|
|
|
self.offset = offset
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
class LinkEnv:
|
|
|
|
def __init__(self, arch):
|
|
|
|
self.arch = ARCH_DATA[arch]
|
|
|
|
self.sections = [] # list of sections in order of output
|
|
|
|
self.literal_sections = [] # list of literal sections (xtensa only)
|
|
|
|
self.known_syms = {} # dict of symbols that are defined
|
|
|
|
self.unresolved_syms = [] # list of unresolved symbols
|
|
|
|
self.mpy_relocs = [] # list of relocations needed in the output .mpy file
|
|
|
|
|
|
|
|
def check_arch(self, arch_name):
|
|
|
|
if arch_name != self.arch.name:
|
|
|
|
raise LinkError("incompatible arch")
|
|
|
|
|
|
|
|
def print_sections(self):
|
|
|
|
log(LOG_LEVEL_2, "sections:")
|
|
|
|
for sec in self.sections:
|
|
|
|
log(LOG_LEVEL_2, " {:08x} {} size={}".format(sec.addr, sec.name, len(sec.data)))
|
|
|
|
|
|
|
|
def find_addr(self, name):
|
|
|
|
if name in self.known_syms:
|
|
|
|
s = self.known_syms[name]
|
|
|
|
return s.section.addr + s["st_value"]
|
|
|
|
raise LinkError("unknown symbol: {}".format(name))
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def build_got_generic(env):
|
|
|
|
env.got_entries = {}
|
|
|
|
for sec in env.sections:
|
|
|
|
for r in sec.reloc:
|
|
|
|
s = r.sym
|
|
|
|
if not (
|
|
|
|
s.entry["st_info"]["bind"] == "STB_GLOBAL"
|
|
|
|
and r["r_info_type"] in env.arch.arch_got
|
|
|
|
):
|
|
|
|
continue
|
|
|
|
s_type = s.entry["st_info"]["type"]
|
|
|
|
assert s_type in ("STT_NOTYPE", "STT_FUNC", "STT_OBJECT"), s_type
|
|
|
|
assert s.name
|
|
|
|
if s.name in env.got_entries:
|
|
|
|
continue
|
|
|
|
env.got_entries[s.name] = GOTEntry(s.name, s)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def build_got_xtensa(env):
|
|
|
|
env.got_entries = {}
|
|
|
|
env.lit_entries = {}
|
|
|
|
env.xt_literals = {}
|
|
|
|
|
|
|
|
# Extract the values from the literal table
|
|
|
|
for sec in env.literal_sections:
|
|
|
|
assert len(sec.data) % env.arch.word_size == 0
|
|
|
|
|
|
|
|
# Look through literal relocations to find any global pointers that should be GOT entries
|
|
|
|
for r in sec.reloc:
|
|
|
|
s = r.sym
|
|
|
|
s_type = s.entry["st_info"]["type"]
|
|
|
|
assert s_type in ("STT_NOTYPE", "STT_FUNC", "STT_OBJECT", "STT_SECTION"), s_type
|
|
|
|
assert r["r_info_type"] in env.arch.arch_got
|
|
|
|
assert r["r_offset"] % env.arch.word_size == 0
|
|
|
|
# This entry is a global pointer
|
|
|
|
existing = struct.unpack_from("<I", sec.data, r["r_offset"])[0]
|
|
|
|
if s_type == "STT_SECTION":
|
|
|
|
assert r["r_addend"] == 0
|
|
|
|
name = "{}+0x{:x}".format(s.section.name, existing)
|
|
|
|
else:
|
|
|
|
assert existing == 0
|
|
|
|
name = s.name
|
|
|
|
if r["r_addend"] != 0:
|
|
|
|
name = "{}+0x{:x}".format(name, r["r_addend"])
|
|
|
|
idx = "{}+0x{:x}".format(sec.filename, r["r_offset"])
|
|
|
|
env.xt_literals[idx] = name
|
|
|
|
if name in env.got_entries:
|
|
|
|
# Deduplicate GOT entries
|
|
|
|
continue
|
|
|
|
env.got_entries[name] = GOTEntry(name, s, existing)
|
|
|
|
|
|
|
|
# Go through all literal entries finding those that aren't global pointers so must be actual literals
|
|
|
|
for i in range(0, len(sec.data), env.arch.word_size):
|
|
|
|
idx = "{}+0x{:x}".format(sec.filename, i)
|
|
|
|
if idx not in env.xt_literals:
|
|
|
|
# This entry is an actual literal
|
|
|
|
value = struct.unpack_from("<I", sec.data, i)[0]
|
|
|
|
env.xt_literals[idx] = value
|
|
|
|
if value in env.lit_entries:
|
|
|
|
# Deduplicate literals
|
|
|
|
continue
|
|
|
|
env.lit_entries[value] = LiteralEntry(
|
|
|
|
value, len(env.lit_entries) * env.arch.word_size
|
|
|
|
)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def populate_got(env):
|
|
|
|
# Compute GOT destination addresses
|
|
|
|
for got_entry in env.got_entries.values():
|
|
|
|
sym = got_entry.sym
|
|
|
|
if hasattr(sym, "resolved"):
|
|
|
|
sym = sym.resolved
|
|
|
|
sec = sym.section
|
|
|
|
addr = sym["st_value"]
|
|
|
|
got_entry.sec_name = sec.name
|
|
|
|
got_entry.link_addr += sec.addr + addr
|
|
|
|
|
|
|
|
# Get sorted GOT, sorted by external, text, rodata, bss so relocations can be combined
|
|
|
|
got_list = sorted(
|
|
|
|
env.got_entries.values(),
|
|
|
|
key=lambda g: g.isexternal() + 2 * g.istext() + 3 * g.isrodata() + 4 * g.isbss(),
|
|
|
|
)
|
|
|
|
|
|
|
|
# Layout and populate the GOT
|
|
|
|
offset = 0
|
|
|
|
for got_entry in got_list:
|
|
|
|
got_entry.offset = offset
|
|
|
|
offset += env.arch.word_size
|
|
|
|
o = env.got_section.addr + got_entry.offset
|
|
|
|
env.full_text[o : o + env.arch.word_size] = got_entry.link_addr.to_bytes(
|
|
|
|
env.arch.word_size, "little"
|
|
|
|
)
|
|
|
|
|
|
|
|
# Create a relocation for each GOT entry
|
|
|
|
for got_entry in got_list:
|
2022-05-26 01:51:29 +01:00
|
|
|
if got_entry.name in ("mp_native_qstr_table", "mp_native_obj_table", "mp_fun_table"):
|
|
|
|
dest = got_entry.name
|
2019-08-28 08:03:08 +01:00
|
|
|
elif got_entry.name.startswith("mp_fun_table+0x"):
|
|
|
|
dest = int(got_entry.name.split("+")[1], 16) // env.arch.word_size
|
2022-06-10 07:38:20 +01:00
|
|
|
elif got_entry.sec_name == ".external.mp_fun_table":
|
|
|
|
dest = got_entry.sym.mp_fun_table_offset
|
2019-08-28 08:03:08 +01:00
|
|
|
elif got_entry.sec_name.startswith(".text"):
|
|
|
|
dest = ".text"
|
|
|
|
elif got_entry.sec_name.startswith(".rodata"):
|
|
|
|
dest = ".rodata"
|
|
|
|
elif got_entry.sec_name.startswith(".data.rel.ro"):
|
|
|
|
dest = ".data.rel.ro"
|
|
|
|
elif got_entry.sec_name.startswith(".bss"):
|
|
|
|
dest = ".bss"
|
|
|
|
else:
|
|
|
|
assert 0, (got_entry.name, got_entry.sec_name)
|
|
|
|
env.mpy_relocs.append((".text", env.got_section.addr + got_entry.offset, dest))
|
|
|
|
|
|
|
|
# Print out the final GOT
|
|
|
|
log(LOG_LEVEL_2, "GOT: {:08x}".format(env.got_section.addr))
|
|
|
|
for g in got_list:
|
|
|
|
log(
|
|
|
|
LOG_LEVEL_2,
|
|
|
|
" {:08x} {} -> {}+{:08x}".format(g.offset, g.name, g.sec_name, g.link_addr),
|
|
|
|
)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def populate_lit(env):
|
|
|
|
log(LOG_LEVEL_2, "LIT: {:08x}".format(env.lit_section.addr))
|
|
|
|
for lit_entry in env.lit_entries.values():
|
|
|
|
value = lit_entry.value
|
|
|
|
log(LOG_LEVEL_2, " {:08x} = {:08x}".format(lit_entry.offset, value))
|
|
|
|
o = env.lit_section.addr + lit_entry.offset
|
|
|
|
env.full_text[o : o + env.arch.word_size] = value.to_bytes(env.arch.word_size, "little")
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def do_relocation_text(env, text_addr, r):
|
|
|
|
# Extract relevant info about symbol that's being relocated
|
|
|
|
s = r.sym
|
|
|
|
s_bind = s.entry["st_info"]["bind"]
|
|
|
|
s_type = s.entry["st_info"]["type"]
|
|
|
|
r_offset = r["r_offset"] + text_addr
|
|
|
|
r_info_type = r["r_info_type"]
|
|
|
|
try:
|
|
|
|
# only for RELA sections
|
|
|
|
r_addend = r["r_addend"]
|
|
|
|
except KeyError:
|
|
|
|
r_addend = 0
|
|
|
|
|
|
|
|
# Default relocation type and name for logging
|
|
|
|
reloc_type = "le32"
|
|
|
|
log_name = None
|
|
|
|
|
|
|
|
if (
|
|
|
|
env.arch.name == "EM_386"
|
|
|
|
and r_info_type in (R_386_PC32, R_386_PLT32)
|
|
|
|
or env.arch.name == "EM_X86_64"
|
|
|
|
and r_info_type in (R_X86_64_PC32, R_X86_64_PLT32)
|
|
|
|
or env.arch.name == "EM_ARM"
|
|
|
|
and r_info_type in (R_ARM_REL32, R_ARM_THM_CALL, R_ARM_THM_JUMP24)
|
|
|
|
or s_bind == "STB_LOCAL"
|
|
|
|
and env.arch.name == "EM_XTENSA"
|
|
|
|
and r_info_type == R_XTENSA_32 # not GOT
|
|
|
|
):
|
|
|
|
# Standard relocation to fixed location within text/rodata
|
|
|
|
if hasattr(s, "resolved"):
|
|
|
|
s = s.resolved
|
|
|
|
|
|
|
|
sec = s.section
|
|
|
|
|
|
|
|
if env.arch.separate_rodata and sec.name.startswith(".rodata"):
|
|
|
|
raise LinkError("fixed relocation to rodata with rodata referenced via GOT")
|
|
|
|
|
|
|
|
if sec.name.startswith(".bss"):
|
|
|
|
raise LinkError(
|
|
|
|
"{}: fixed relocation to bss (bss variables can't be static)".format(s.filename)
|
2020-02-27 04:36:53 +00:00
|
|
|
)
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
if sec.name.startswith(".external"):
|
|
|
|
raise LinkError(
|
|
|
|
"{}: fixed relocation to external symbol: {}".format(s.filename, s.name)
|
2020-02-27 04:36:53 +00:00
|
|
|
)
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
addr = sec.addr + s["st_value"]
|
|
|
|
reloc = addr - r_offset + r_addend
|
|
|
|
|
|
|
|
if r_info_type in (R_ARM_THM_CALL, R_ARM_THM_JUMP24):
|
|
|
|
# Both relocations have the same bit pattern to rewrite:
|
|
|
|
# R_ARM_THM_CALL: bl
|
|
|
|
# R_ARM_THM_JUMP24: b.w
|
|
|
|
reloc_type = "thumb_b"
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
elif (
|
|
|
|
env.arch.name == "EM_386"
|
|
|
|
and r_info_type == R_386_GOTPC
|
|
|
|
or env.arch.name == "EM_ARM"
|
|
|
|
and r_info_type == R_ARM_BASE_PREL
|
|
|
|
):
|
|
|
|
# Relocation to GOT address itself
|
|
|
|
assert s.name == "_GLOBAL_OFFSET_TABLE_"
|
|
|
|
addr = env.got_section.addr
|
|
|
|
reloc = addr - r_offset + r_addend
|
|
|
|
|
|
|
|
elif (
|
|
|
|
env.arch.name == "EM_386"
|
|
|
|
and r_info_type in (R_386_GOT32, R_386_GOT32X)
|
|
|
|
or env.arch.name == "EM_ARM"
|
|
|
|
and r_info_type == R_ARM_GOT_BREL
|
|
|
|
):
|
|
|
|
# Relcation pointing to GOT
|
|
|
|
reloc = addr = env.got_entries[s.name].offset
|
|
|
|
|
2020-11-14 07:54:26 +00:00
|
|
|
elif env.arch.name == "EM_X86_64" and r_info_type in (
|
|
|
|
R_X86_64_GOTPCREL,
|
|
|
|
R_X86_64_REX_GOTPCRELX,
|
|
|
|
):
|
2019-08-28 08:03:08 +01:00
|
|
|
# Relcation pointing to GOT
|
|
|
|
got_entry = env.got_entries[s.name]
|
|
|
|
addr = env.got_section.addr + got_entry.offset
|
|
|
|
reloc = addr - r_offset + r_addend
|
|
|
|
|
|
|
|
elif env.arch.name == "EM_386" and r_info_type == R_386_GOTOFF:
|
|
|
|
# Relocation relative to GOT
|
|
|
|
addr = s.section.addr + s["st_value"]
|
|
|
|
reloc = addr - env.got_section.addr + r_addend
|
|
|
|
|
|
|
|
elif env.arch.name == "EM_XTENSA" and r_info_type == R_XTENSA_SLOT0_OP:
|
|
|
|
# Relocation pointing to GOT, xtensa specific
|
|
|
|
sec = s.section
|
|
|
|
if sec.name.startswith(".text"):
|
|
|
|
# it looks like R_XTENSA_SLOT0_OP into .text is already correctly relocated
|
|
|
|
return
|
|
|
|
assert sec.name.startswith(".literal"), sec.name
|
|
|
|
lit_idx = "{}+0x{:x}".format(sec.filename, r_addend)
|
|
|
|
lit_ptr = env.xt_literals[lit_idx]
|
|
|
|
if isinstance(lit_ptr, str):
|
|
|
|
addr = env.got_section.addr + env.got_entries[lit_ptr].offset
|
|
|
|
log_name = "GOT {}".format(lit_ptr)
|
|
|
|
else:
|
|
|
|
addr = env.lit_section.addr + env.lit_entries[lit_ptr].offset
|
|
|
|
log_name = "LIT"
|
|
|
|
reloc = addr - r_offset
|
|
|
|
reloc_type = "xtensa_l32r"
|
|
|
|
|
2022-06-09 02:57:22 +01:00
|
|
|
elif env.arch.name == "EM_XTENSA" and r_info_type in (R_XTENSA_DIFF32, R_XTENSA_PDIFF32):
|
2019-08-28 08:03:08 +01:00
|
|
|
if s.section.name.startswith(".text"):
|
2022-06-09 02:57:22 +01:00
|
|
|
# it looks like R_XTENSA_[P]DIFF32 into .text is already correctly relocated
|
2019-08-28 08:03:08 +01:00
|
|
|
return
|
|
|
|
assert 0
|
|
|
|
|
|
|
|
else:
|
|
|
|
# Unknown/unsupported relocation
|
|
|
|
assert 0, r_info_type
|
|
|
|
|
|
|
|
# Write relocation
|
|
|
|
if reloc_type == "le32":
|
|
|
|
(existing,) = struct.unpack_from("<I", env.full_text, r_offset)
|
|
|
|
struct.pack_into("<I", env.full_text, r_offset, (existing + reloc) & 0xFFFFFFFF)
|
|
|
|
elif reloc_type == "thumb_b":
|
|
|
|
b_h, b_l = struct.unpack_from("<HH", env.full_text, r_offset)
|
|
|
|
existing = (b_h & 0x7FF) << 12 | (b_l & 0x7FF) << 1
|
|
|
|
if existing >= 0x400000: # 2's complement
|
|
|
|
existing -= 0x800000
|
|
|
|
new = existing + reloc
|
|
|
|
b_h = (b_h & 0xF800) | (new >> 12) & 0x7FF
|
|
|
|
b_l = (b_l & 0xF800) | (new >> 1) & 0x7FF
|
|
|
|
struct.pack_into("<HH", env.full_text, r_offset, b_h, b_l)
|
|
|
|
elif reloc_type == "xtensa_l32r":
|
|
|
|
l32r = unpack_u24le(env.full_text, r_offset)
|
|
|
|
assert l32r & 0xF == 1 # RI16 encoded l32r
|
|
|
|
l32r_imm16 = l32r >> 8
|
|
|
|
l32r_imm16 = (l32r_imm16 + reloc >> 2) & 0xFFFF
|
|
|
|
l32r = l32r & 0xFF | l32r_imm16 << 8
|
|
|
|
pack_u24le(env.full_text, r_offset, l32r)
|
|
|
|
else:
|
|
|
|
assert 0, reloc_type
|
|
|
|
|
|
|
|
# Log information about relocation
|
|
|
|
if log_name is None:
|
|
|
|
if s_type == "STT_SECTION":
|
|
|
|
log_name = s.section.name
|
|
|
|
else:
|
|
|
|
log_name = s.name
|
|
|
|
log(LOG_LEVEL_3, " {:08x} {} -> {:08x}".format(r_offset, log_name, addr))
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def do_relocation_data(env, text_addr, r):
|
|
|
|
s = r.sym
|
|
|
|
s_type = s.entry["st_info"]["type"]
|
|
|
|
r_offset = r["r_offset"] + text_addr
|
|
|
|
r_info_type = r["r_info_type"]
|
|
|
|
try:
|
|
|
|
# only for RELA sections
|
|
|
|
r_addend = r["r_addend"]
|
|
|
|
except KeyError:
|
|
|
|
r_addend = 0
|
|
|
|
|
|
|
|
if (
|
|
|
|
env.arch.name == "EM_386"
|
|
|
|
and r_info_type == R_386_32
|
|
|
|
or env.arch.name == "EM_X86_64"
|
|
|
|
and r_info_type == R_X86_64_64
|
|
|
|
or env.arch.name == "EM_ARM"
|
|
|
|
and r_info_type == R_ARM_ABS32
|
|
|
|
or env.arch.name == "EM_XTENSA"
|
|
|
|
and r_info_type == R_XTENSA_32
|
|
|
|
):
|
|
|
|
# Relocation in data.rel.ro to internal/external symbol
|
|
|
|
if env.arch.word_size == 4:
|
|
|
|
struct_type = "<I"
|
|
|
|
elif env.arch.word_size == 8:
|
|
|
|
struct_type = "<Q"
|
|
|
|
sec = s.section
|
|
|
|
assert r_offset % env.arch.word_size == 0
|
|
|
|
addr = sec.addr + s["st_value"] + r_addend
|
|
|
|
if s_type == "STT_SECTION":
|
|
|
|
log_name = sec.name
|
|
|
|
else:
|
|
|
|
log_name = s.name
|
|
|
|
log(LOG_LEVEL_3, " {:08x} -> {} {:08x}".format(r_offset, log_name, addr))
|
|
|
|
if env.arch.separate_rodata:
|
|
|
|
data = env.full_rodata
|
|
|
|
else:
|
|
|
|
data = env.full_text
|
|
|
|
(existing,) = struct.unpack_from(struct_type, data, r_offset)
|
|
|
|
if sec.name.startswith((".text", ".rodata", ".data.rel.ro", ".bss")):
|
|
|
|
struct.pack_into(struct_type, data, r_offset, existing + addr)
|
|
|
|
kind = sec.name
|
|
|
|
elif sec.name == ".external.mp_fun_table":
|
|
|
|
assert addr == 0
|
|
|
|
kind = s.mp_fun_table_offset
|
|
|
|
else:
|
|
|
|
assert 0, sec.name
|
|
|
|
if env.arch.separate_rodata:
|
|
|
|
base = ".rodata"
|
|
|
|
else:
|
|
|
|
base = ".text"
|
|
|
|
env.mpy_relocs.append((base, r_offset, kind))
|
|
|
|
|
|
|
|
else:
|
|
|
|
# Unknown/unsupported relocation
|
|
|
|
assert 0, r_info_type
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def load_object_file(env, felf):
|
|
|
|
with open(felf, "rb") as f:
|
|
|
|
elf = elffile.ELFFile(f)
|
|
|
|
env.check_arch(elf["e_machine"])
|
|
|
|
|
|
|
|
# Get symbol table
|
|
|
|
symtab = list(elf.get_section_by_name(".symtab").iter_symbols())
|
|
|
|
|
|
|
|
# Load needed sections from ELF file
|
|
|
|
sections_shndx = {} # maps elf shndx to Section object
|
|
|
|
for idx, s in enumerate(elf.iter_sections()):
|
|
|
|
if s.header.sh_type in ("SHT_PROGBITS", "SHT_NOBITS"):
|
|
|
|
if s.data_size == 0:
|
|
|
|
# Ignore empty sections
|
|
|
|
pass
|
|
|
|
elif s.name.startswith((".literal", ".text", ".rodata", ".data.rel.ro", ".bss")):
|
|
|
|
sec = Section.from_elfsec(s, felf)
|
|
|
|
sections_shndx[idx] = sec
|
|
|
|
if s.name.startswith(".literal"):
|
|
|
|
env.literal_sections.append(sec)
|
|
|
|
else:
|
|
|
|
env.sections.append(sec)
|
|
|
|
elif s.name.startswith(".data"):
|
|
|
|
raise LinkError("{}: {} non-empty".format(felf, s.name))
|
|
|
|
else:
|
|
|
|
# Ignore section
|
|
|
|
pass
|
|
|
|
elif s.header.sh_type in ("SHT_REL", "SHT_RELA"):
|
|
|
|
shndx = s.header.sh_info
|
|
|
|
if shndx in sections_shndx:
|
|
|
|
sec = sections_shndx[shndx]
|
|
|
|
sec.reloc_name = s.name
|
|
|
|
sec.reloc = list(s.iter_relocations())
|
|
|
|
for r in sec.reloc:
|
|
|
|
r.sym = symtab[r["r_info_sym"]]
|
|
|
|
|
|
|
|
# Link symbols to their sections, and update known and unresolved symbols
|
|
|
|
for sym in symtab:
|
|
|
|
sym.filename = felf
|
|
|
|
shndx = sym.entry["st_shndx"]
|
|
|
|
if shndx in sections_shndx:
|
|
|
|
# Symbol with associated section
|
|
|
|
sym.section = sections_shndx[shndx]
|
|
|
|
if sym["st_info"]["bind"] == "STB_GLOBAL":
|
|
|
|
# Defined global symbol
|
|
|
|
if sym.name in env.known_syms and not sym.name.startswith(
|
|
|
|
"__x86.get_pc_thunk."
|
|
|
|
):
|
|
|
|
raise LinkError("duplicate symbol: {}".format(sym.name))
|
|
|
|
env.known_syms[sym.name] = sym
|
|
|
|
elif sym.entry["st_shndx"] == "SHN_UNDEF" and sym["st_info"]["bind"] == "STB_GLOBAL":
|
|
|
|
# Undefined global symbol, needs resolving
|
|
|
|
env.unresolved_syms.append(sym)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def link_objects(env, native_qstr_vals_len, native_qstr_objs_len):
|
|
|
|
# Build GOT information
|
|
|
|
if env.arch.name == "EM_XTENSA":
|
|
|
|
build_got_xtensa(env)
|
|
|
|
else:
|
|
|
|
build_got_generic(env)
|
|
|
|
|
|
|
|
# Creat GOT section
|
|
|
|
got_size = len(env.got_entries) * env.arch.word_size
|
|
|
|
env.got_section = Section("GOT", bytearray(got_size), env.arch.word_size)
|
|
|
|
if env.arch.name == "EM_XTENSA":
|
|
|
|
env.sections.insert(0, env.got_section)
|
|
|
|
else:
|
|
|
|
env.sections.append(env.got_section)
|
|
|
|
|
|
|
|
# Create optional literal section
|
|
|
|
if env.arch.name == "EM_XTENSA":
|
|
|
|
lit_size = len(env.lit_entries) * env.arch.word_size
|
|
|
|
env.lit_section = Section("LIT", bytearray(lit_size), env.arch.word_size)
|
|
|
|
env.sections.insert(1, env.lit_section)
|
|
|
|
|
2022-05-26 01:51:29 +01:00
|
|
|
# Create section to contain mp_native_qstr_table
|
|
|
|
env.qstr_table_section = Section(
|
|
|
|
".external.qstr_table",
|
2019-08-28 08:03:08 +01:00
|
|
|
bytearray(native_qstr_vals_len * env.arch.qstr_entry_size),
|
|
|
|
env.arch.qstr_entry_size,
|
|
|
|
)
|
|
|
|
|
2022-05-26 01:51:29 +01:00
|
|
|
# Create section to contain mp_native_obj_table
|
|
|
|
env.obj_table_section = Section(
|
|
|
|
".external.obj_table",
|
|
|
|
bytearray(native_qstr_objs_len * env.arch.word_size),
|
|
|
|
env.arch.word_size,
|
2019-08-28 08:03:08 +01:00
|
|
|
)
|
|
|
|
|
|
|
|
# Resolve unknown symbols
|
|
|
|
mp_fun_table_sec = Section(".external.mp_fun_table", b"", 0)
|
2019-11-30 12:03:09 +00:00
|
|
|
fun_table = {
|
|
|
|
key: 67 + idx
|
2019-08-28 08:03:08 +01:00
|
|
|
for idx, key in enumerate(
|
|
|
|
[
|
|
|
|
"mp_type_type",
|
|
|
|
"mp_type_str",
|
|
|
|
"mp_type_list",
|
|
|
|
"mp_type_dict",
|
|
|
|
"mp_type_fun_builtin_0",
|
|
|
|
"mp_type_fun_builtin_1",
|
|
|
|
"mp_type_fun_builtin_2",
|
|
|
|
"mp_type_fun_builtin_3",
|
|
|
|
"mp_type_fun_builtin_var",
|
|
|
|
"mp_stream_read_obj",
|
|
|
|
"mp_stream_readinto_obj",
|
|
|
|
"mp_stream_unbuffered_readline_obj",
|
|
|
|
"mp_stream_write_obj",
|
|
|
|
]
|
|
|
|
)
|
|
|
|
}
|
|
|
|
for sym in env.unresolved_syms:
|
|
|
|
assert sym["st_value"] == 0
|
|
|
|
if sym.name == "_GLOBAL_OFFSET_TABLE_":
|
|
|
|
pass
|
|
|
|
elif sym.name == "mp_fun_table":
|
|
|
|
sym.section = Section(".external", b"", 0)
|
2022-05-26 01:51:29 +01:00
|
|
|
elif sym.name == "mp_native_qstr_table":
|
|
|
|
sym.section = env.qstr_table_section
|
|
|
|
elif sym.name == "mp_native_obj_table":
|
|
|
|
sym.section = env.obj_table_section
|
2019-08-28 08:03:08 +01:00
|
|
|
elif sym.name in env.known_syms:
|
|
|
|
sym.resolved = env.known_syms[sym.name]
|
|
|
|
else:
|
|
|
|
if sym.name in fun_table:
|
|
|
|
sym.section = mp_fun_table_sec
|
|
|
|
sym.mp_fun_table_offset = fun_table[sym.name]
|
|
|
|
else:
|
|
|
|
raise LinkError("{}: undefined symbol: {}".format(sym.filename, sym.name))
|
|
|
|
|
|
|
|
# Align sections, assign their addresses, and create full_text
|
|
|
|
env.full_text = bytearray(env.arch.asm_jump(8)) # dummy, to be filled in later
|
|
|
|
env.full_rodata = bytearray(0)
|
|
|
|
env.full_bss = bytearray(0)
|
|
|
|
for sec in env.sections:
|
|
|
|
if env.arch.separate_rodata and sec.name.startswith((".rodata", ".data.rel.ro")):
|
|
|
|
data = env.full_rodata
|
|
|
|
elif sec.name.startswith(".bss"):
|
|
|
|
data = env.full_bss
|
|
|
|
else:
|
|
|
|
data = env.full_text
|
|
|
|
sec.addr = align_to(len(data), sec.alignment)
|
|
|
|
data.extend(b"\x00" * (sec.addr - len(data)))
|
|
|
|
data.extend(sec.data)
|
|
|
|
|
|
|
|
env.print_sections()
|
|
|
|
|
|
|
|
populate_got(env)
|
|
|
|
if env.arch.name == "EM_XTENSA":
|
|
|
|
populate_lit(env)
|
|
|
|
|
|
|
|
# Fill in relocations
|
|
|
|
for sec in env.sections:
|
|
|
|
if not sec.reloc:
|
|
|
|
continue
|
|
|
|
log(
|
|
|
|
LOG_LEVEL_3,
|
|
|
|
"{}: {} relocations via {}:".format(sec.filename, sec.name, sec.reloc_name),
|
|
|
|
)
|
|
|
|
for r in sec.reloc:
|
|
|
|
if sec.name.startswith((".text", ".rodata")):
|
|
|
|
do_relocation_text(env, sec.addr, r)
|
|
|
|
elif sec.name.startswith(".data.rel.ro"):
|
|
|
|
do_relocation_data(env, sec.addr, r)
|
|
|
|
else:
|
|
|
|
assert 0, sec.name
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
################################################################################
|
|
|
|
# .mpy output
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
class MPYOutput:
|
|
|
|
def open(self, fname):
|
|
|
|
self.f = open(fname, "wb")
|
|
|
|
self.prev_base = -1
|
|
|
|
self.prev_offset = -1
|
|
|
|
|
|
|
|
def close(self):
|
|
|
|
self.f.close()
|
|
|
|
|
|
|
|
def write_bytes(self, buf):
|
|
|
|
self.f.write(buf)
|
|
|
|
|
|
|
|
def write_uint(self, val):
|
|
|
|
b = bytearray()
|
|
|
|
b.insert(0, val & 0x7F)
|
|
|
|
val >>= 7
|
|
|
|
while val:
|
|
|
|
b.insert(0, 0x80 | (val & 0x7F))
|
|
|
|
val >>= 7
|
|
|
|
self.write_bytes(b)
|
|
|
|
|
|
|
|
def write_qstr(self, s):
|
|
|
|
if s in qstrutil.static_qstr_list:
|
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
|
|
|
self.write_uint((qstrutil.static_qstr_list.index(s) + 1) << 1 | 1)
|
2019-08-28 08:03:08 +01:00
|
|
|
else:
|
|
|
|
s = bytes(s, "ascii")
|
|
|
|
self.write_uint(len(s) << 1)
|
|
|
|
self.write_bytes(s)
|
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
|
|
|
self.write_bytes(b"\x00")
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def write_reloc(self, base, offset, dest, n):
|
|
|
|
need_offset = not (base == self.prev_base and offset == self.prev_offset + 1)
|
|
|
|
self.prev_offset = offset + n - 1
|
|
|
|
if dest <= 2:
|
|
|
|
dest = (dest << 1) | (n > 1)
|
|
|
|
else:
|
|
|
|
assert 6 <= dest <= 127
|
|
|
|
assert n == 1
|
|
|
|
dest = dest << 1 | need_offset
|
|
|
|
assert 0 <= dest <= 0xFE, dest
|
|
|
|
self.write_bytes(bytes([dest]))
|
|
|
|
if need_offset:
|
|
|
|
if base == ".text":
|
|
|
|
base = 0
|
|
|
|
elif base == ".rodata":
|
|
|
|
base = 1
|
|
|
|
self.write_uint(offset << 1 | base)
|
|
|
|
if n > 1:
|
|
|
|
self.write_uint(n)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def build_mpy(env, entry_offset, fmpy, native_qstr_vals, native_qstr_objs):
|
|
|
|
# Write jump instruction to start of text
|
|
|
|
jump = env.arch.asm_jump(entry_offset)
|
|
|
|
env.full_text[: len(jump)] = jump
|
|
|
|
|
|
|
|
log(LOG_LEVEL_1, "arch: {}".format(env.arch.name))
|
|
|
|
log(LOG_LEVEL_1, "text size: {}".format(len(env.full_text)))
|
|
|
|
if len(env.full_rodata):
|
|
|
|
log(LOG_LEVEL_1, "rodata size: {}".format(len(env.full_rodata)))
|
|
|
|
log(LOG_LEVEL_1, "bss size: {}".format(len(env.full_bss)))
|
|
|
|
log(LOG_LEVEL_1, "GOT entries: {}".format(len(env.got_entries)))
|
|
|
|
|
|
|
|
# xxd(env.full_text)
|
|
|
|
|
|
|
|
out = MPYOutput()
|
|
|
|
out.open(fmpy)
|
|
|
|
|
|
|
|
# MPY: header
|
2022-09-17 14:57:12 +01:00
|
|
|
out.write_bytes(
|
|
|
|
bytearray(
|
|
|
|
[ord("M"), MPY_VERSION, env.arch.mpy_feature | MPY_SUB_VERSION, MP_SMALL_INT_BITS]
|
|
|
|
)
|
|
|
|
)
|
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
|
|
|
|
|
|
|
# MPY: n_qstr
|
2022-05-26 01:51:29 +01:00
|
|
|
out.write_uint(1 + len(native_qstr_vals))
|
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
|
|
|
|
|
|
|
# MPY: n_obj
|
2022-05-26 01:51:29 +01:00
|
|
|
out.write_uint(len(native_qstr_objs))
|
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
|
|
|
|
|
|
|
# MPY: qstr table
|
|
|
|
out.write_qstr(fmpy) # filename
|
2022-05-26 01:51:29 +01:00
|
|
|
for q in native_qstr_vals:
|
|
|
|
out.write_qstr(q)
|
|
|
|
|
|
|
|
# MPY: object table
|
|
|
|
for q in native_qstr_objs:
|
|
|
|
out.write_bytes(bytearray([MP_PERSISTENT_OBJ_STR]))
|
|
|
|
out.write_uint(len(q))
|
|
|
|
out.write_bytes(bytes(q, "utf8") + b"\x00")
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
# MPY: kind/len
|
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
|
|
|
out.write_uint(len(env.full_text) << 3 | (MP_CODE_NATIVE_VIPER - MP_CODE_BYTECODE))
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
# MPY: machine code
|
|
|
|
out.write_bytes(env.full_text)
|
|
|
|
|
|
|
|
# MPY: scope_flags
|
|
|
|
scope_flags = MP_SCOPE_FLAG_VIPERRELOC
|
|
|
|
if len(env.full_rodata):
|
|
|
|
scope_flags |= MP_SCOPE_FLAG_VIPERRODATA
|
|
|
|
if len(env.full_bss):
|
|
|
|
scope_flags |= MP_SCOPE_FLAG_VIPERBSS
|
|
|
|
out.write_uint(scope_flags)
|
|
|
|
|
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
|
|
|
# MPY: bss and/or rodata
|
2019-08-28 08:03:08 +01:00
|
|
|
if len(env.full_rodata):
|
|
|
|
rodata_const_table_idx = 1
|
|
|
|
out.write_uint(len(env.full_rodata))
|
|
|
|
if len(env.full_bss):
|
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
|
|
|
bss_const_table_idx = 2
|
2019-08-28 08:03:08 +01:00
|
|
|
out.write_uint(len(env.full_bss))
|
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 len(env.full_rodata):
|
|
|
|
out.write_bytes(env.full_rodata)
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
# MPY: relocation information
|
|
|
|
prev_kind = None
|
2023-08-09 09:22:01 +01:00
|
|
|
prev_base = None
|
|
|
|
prev_offset = None
|
|
|
|
prev_n = None
|
2019-08-28 08:03:08 +01:00
|
|
|
for base, addr, kind in env.mpy_relocs:
|
|
|
|
if isinstance(kind, str) and kind.startswith(".text"):
|
|
|
|
kind = 0
|
2023-07-30 14:40:39 +01:00
|
|
|
elif isinstance(kind, str) and kind.startswith((".rodata", ".data.rel.ro")):
|
2019-08-28 08:03:08 +01:00
|
|
|
if env.arch.separate_rodata:
|
|
|
|
kind = rodata_const_table_idx
|
|
|
|
else:
|
|
|
|
kind = 0
|
|
|
|
elif isinstance(kind, str) and kind.startswith(".bss"):
|
|
|
|
kind = bss_const_table_idx
|
2022-05-26 01:51:29 +01:00
|
|
|
elif kind == "mp_native_qstr_table":
|
2019-08-28 08:03:08 +01:00
|
|
|
kind = 6
|
2022-05-26 01:51:29 +01:00
|
|
|
elif kind == "mp_native_obj_table":
|
|
|
|
kind = 7
|
|
|
|
elif kind == "mp_fun_table":
|
|
|
|
kind = 8
|
2019-08-28 08:03:08 +01:00
|
|
|
else:
|
2022-05-26 01:51:29 +01:00
|
|
|
kind = 9 + kind
|
2019-08-28 08:03:08 +01:00
|
|
|
assert addr % env.arch.word_size == 0, addr
|
|
|
|
offset = addr // env.arch.word_size
|
|
|
|
if kind == prev_kind and base == prev_base and offset == prev_offset + 1:
|
|
|
|
prev_n += 1
|
|
|
|
prev_offset += 1
|
|
|
|
else:
|
|
|
|
if prev_kind is not None:
|
|
|
|
out.write_reloc(prev_base, prev_offset - prev_n + 1, prev_kind, prev_n)
|
|
|
|
prev_kind = kind
|
|
|
|
prev_base = base
|
|
|
|
prev_offset = offset
|
|
|
|
prev_n = 1
|
|
|
|
if prev_kind is not None:
|
|
|
|
out.write_reloc(prev_base, prev_offset - prev_n + 1, prev_kind, prev_n)
|
|
|
|
|
|
|
|
# MPY: sentinel for end of relocations
|
|
|
|
out.write_bytes(b"\xff")
|
|
|
|
|
|
|
|
out.close()
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
################################################################################
|
|
|
|
# main
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def do_preprocess(args):
|
|
|
|
if args.output is None:
|
|
|
|
assert args.files[0].endswith(".c")
|
|
|
|
args.output = args.files[0][:-1] + "config.h"
|
|
|
|
static_qstrs, qstr_vals, qstr_objs = extract_qstrs(args.files)
|
|
|
|
with open(args.output, "w") as f:
|
|
|
|
print(
|
|
|
|
"#include <stdint.h>\n"
|
|
|
|
"typedef uintptr_t mp_uint_t;\n"
|
|
|
|
"typedef intptr_t mp_int_t;\n"
|
|
|
|
"typedef uintptr_t mp_off_t;",
|
|
|
|
file=f,
|
|
|
|
)
|
|
|
|
for i, q in enumerate(static_qstrs):
|
|
|
|
print("#define %s (%u)" % (q, i + 1), file=f)
|
|
|
|
for i, q in enumerate(sorted(qstr_vals)):
|
2022-05-26 01:51:29 +01:00
|
|
|
print("#define %s (mp_native_qstr_table[%d])" % (q, i + 1), file=f)
|
2019-08-28 08:03:08 +01:00
|
|
|
for i, q in enumerate(sorted(qstr_objs)):
|
|
|
|
print(
|
2022-05-26 01:51:29 +01:00
|
|
|
"#define MP_OBJ_NEW_QSTR_%s ((mp_obj_t)mp_native_obj_table[%d])" % (q, i),
|
2019-08-28 08:03:08 +01:00
|
|
|
file=f,
|
|
|
|
)
|
2022-05-26 01:51:29 +01:00
|
|
|
print("extern const uint16_t mp_native_qstr_table[];", file=f)
|
|
|
|
print("extern const mp_uint_t mp_native_obj_table[];", file=f)
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
|
|
|
|
def do_link(args):
|
|
|
|
if args.output is None:
|
|
|
|
assert args.files[0].endswith(".o")
|
|
|
|
args.output = args.files[0][:-1] + "mpy"
|
|
|
|
native_qstr_vals = []
|
|
|
|
native_qstr_objs = []
|
|
|
|
if args.qstrs is not None:
|
|
|
|
with open(args.qstrs) as f:
|
|
|
|
for l in f:
|
|
|
|
m = re.match(r"#define MP_QSTR_([A-Za-z0-9_]*) \(mp_native_", l)
|
|
|
|
if m:
|
|
|
|
native_qstr_vals.append(m.group(1))
|
|
|
|
else:
|
|
|
|
m = re.match(r"#define MP_OBJ_NEW_QSTR_MP_QSTR_([A-Za-z0-9_]*)", l)
|
|
|
|
if m:
|
|
|
|
native_qstr_objs.append(m.group(1))
|
|
|
|
log(LOG_LEVEL_2, "qstr vals: " + ", ".join(native_qstr_vals))
|
|
|
|
log(LOG_LEVEL_2, "qstr objs: " + ", ".join(native_qstr_objs))
|
|
|
|
env = LinkEnv(args.arch)
|
|
|
|
try:
|
|
|
|
for file in args.files:
|
|
|
|
load_object_file(env, file)
|
|
|
|
link_objects(env, len(native_qstr_vals), len(native_qstr_objs))
|
|
|
|
build_mpy(env, env.find_addr("mpy_init"), args.output, native_qstr_vals, native_qstr_objs)
|
|
|
|
except LinkError as er:
|
|
|
|
print("LinkError:", er.args[0])
|
|
|
|
sys.exit(1)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
def main():
|
|
|
|
import argparse
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
cmd_parser = argparse.ArgumentParser(description="Run scripts on the pyboard.")
|
|
|
|
cmd_parser.add_argument(
|
|
|
|
"--verbose", "-v", action="count", default=1, help="increase verbosity"
|
|
|
|
)
|
|
|
|
cmd_parser.add_argument("--arch", default="x64", help="architecture")
|
|
|
|
cmd_parser.add_argument("--preprocess", action="store_true", help="preprocess source files")
|
|
|
|
cmd_parser.add_argument("--qstrs", default=None, help="file defining additional qstrs")
|
|
|
|
cmd_parser.add_argument(
|
|
|
|
"--output", "-o", default=None, help="output .mpy file (default to input with .o->.mpy)"
|
|
|
|
)
|
|
|
|
cmd_parser.add_argument("files", nargs="+", help="input files")
|
|
|
|
args = cmd_parser.parse_args()
|
|
|
|
|
|
|
|
global log_level
|
|
|
|
log_level = args.verbose
|
|
|
|
|
|
|
|
if args.preprocess:
|
|
|
|
do_preprocess(args)
|
|
|
|
else:
|
|
|
|
do_link(args)
|
|
|
|
|
2020-02-27 04:36:53 +00:00
|
|
|
|
2019-08-28 08:03:08 +01:00
|
|
|
if __name__ == "__main__":
|
|
|
|
main()
|