circuitpython/tools/mpy_ld.py

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

1090 lines
36 KiB
Python
Raw Normal View History

#!/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 07:22:47 -04:00
MPY_VERSION = 6
MP_CODE_BYTECODE = 2
MP_CODE_NATIVE_VIPER = 4
MP_NATIVE_ARCH_X86 = 1
MP_NATIVE_ARCH_X64 = 2
MP_NATIVE_ARCH_ARMV6M = 4
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
MP_PERSISTENT_OBJ_STR = 5
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
R_X86_64_GOTPCREL = 9
R_X86_64_REX_GOTPCRELX = 42
R_386_GOT32X = 43
R_XTENSA_PDIFF32 = 59
################################################################################
# Architecture configuration
def asm_jump_x86(entry):
return struct.pack("<BI", 0xE9, entry - 5)
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):
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)
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)
class ArchData:
def __init__(self, name, mpy_feature, word_size, arch_got, asm_jump, *, separate_rodata=False):
self.name = name
self.mpy_feature = mpy_feature
self.qstr_entry_size = 2
self.word_size = word_size
self.arch_got = arch_got
self.asm_jump = asm_jump
self.separate_rodata = separate_rodata
ARCH_DATA = {
"x86": ArchData(
"EM_386",
MP_NATIVE_ARCH_X86 << 2,
4,
(R_386_PC32, R_386_GOT32, R_386_GOT32X),
asm_jump_x86,
),
"x64": ArchData(
"EM_X86_64",
MP_NATIVE_ARCH_X64 << 2,
8,
(R_X86_64_GOTPCREL, R_X86_64_REX_GOTPCRELX),
asm_jump_x86,
),
"armv6m": ArchData(
"EM_ARM",
MP_NATIVE_ARCH_ARMV6M << 2,
4,
(R_ARM_GOT_BREL,),
asm_jump_thumb,
),
"armv7m": ArchData(
"EM_ARM",
MP_NATIVE_ARCH_ARMV7M << 2,
4,
(R_ARM_GOT_BREL,),
asm_jump_thumb2,
),
"armv7emsp": ArchData(
"EM_ARM",
MP_NATIVE_ARCH_ARMV7EMSP << 2,
4,
(R_ARM_GOT_BREL,),
asm_jump_thumb2,
),
"armv7emdp": ArchData(
"EM_ARM",
MP_NATIVE_ARCH_ARMV7EMDP << 2,
4,
(R_ARM_GOT_BREL,),
asm_jump_thumb2,
),
"xtensa": ArchData(
"EM_XTENSA",
MP_NATIVE_ARCH_XTENSA << 2,
4,
(R_XTENSA_32, R_XTENSA_PLT),
asm_jump_xtensa,
),
"xtensawin": ArchData(
"EM_XTENSA",
MP_NATIVE_ARCH_XTENSAWIN << 2,
4,
(R_XTENSA_32, R_XTENSA_PLT),
asm_jump_xtensa,
separate_rodata=True,
),
}
################################################################################
# Helper functions
def align_to(value, align):
return (value + align - 1) & ~(align - 1)
def unpack_u24le(data, offset):
return data[offset] | data[offset + 1] << 8 | data[offset + 2] << 16
def pack_u24le(data, offset, value):
data[offset] = value & 0xFF
data[offset + 1] = value >> 8 & 0xFF
data[offset + 2] = value >> 16 & 0xFF
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()
# Smaller numbers are enabled first
LOG_LEVEL_1 = 1
LOG_LEVEL_2 = 2
LOG_LEVEL_3 = 3
log_level = LOG_LEVEL_1
def log(level, msg):
if level <= log_level:
print(msg)
################################################################################
# Qstr extraction
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
################################################################################
# Linker
class LinkError(Exception):
pass
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)
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")
class LiteralEntry:
def __init__(self, value, offset):
self.value = value
self.offset = offset
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))
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)
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
)
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:
if got_entry.name in ("mp_native_qstr_table", "mp_native_obj_table", "mp_fun_table"):
dest = got_entry.name
elif got_entry.name.startswith("mp_fun_table+0x"):
dest = int(got_entry.name.split("+")[1], 16) // env.arch.word_size
elif got_entry.sec_name == ".external.mp_fun_table":
dest = got_entry.sym.mp_fun_table_offset
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),
)
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")
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_shndx = s.entry["st_shndx"]
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)
)
if sec.name.startswith(".external"):
raise LinkError(
"{}: fixed relocation to external symbol: {}".format(s.filename, s.name)
)
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"
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
elif env.arch.name == "EM_X86_64" and r_info_type in (
R_X86_64_GOTPCREL,
R_X86_64_REX_GOTPCRELX,
):
# 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"
elif env.arch.name == "EM_XTENSA" and r_info_type in (R_XTENSA_DIFF32, R_XTENSA_PDIFF32):
if s.section.name.startswith(".text"):
# it looks like R_XTENSA_[P]DIFF32 into .text is already correctly relocated
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))
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
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)
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)
# Create section to contain mp_native_qstr_table
env.qstr_table_section = Section(
".external.qstr_table",
bytearray(native_qstr_vals_len * env.arch.qstr_entry_size),
env.arch.qstr_entry_size,
)
# 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,
)
# Resolve unknown symbols
mp_fun_table_sec = Section(".external.mp_fun_table", b"", 0)
fun_table = {
key: 67 + idx
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)
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
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
################################################################################
# .mpy output
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 07:22:47 -04:00
self.write_uint((qstrutil.static_qstr_list.index(s) + 1) << 1 | 1)
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 07:22:47 -04:00
self.write_bytes(b"\x00")
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)
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
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 07:22:47 -04:00
out.write_bytes(bytearray([ord("M"), MPY_VERSION, env.arch.mpy_feature, MP_SMALL_INT_BITS]))
# MPY: n_qstr
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 07:22:47 -04:00
# MPY: n_obj
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 07:22:47 -04:00
# MPY: qstr table
out.write_qstr(fmpy) # filename
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")
# 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 07:22:47 -04:00
out.write_uint(len(env.full_text) << 3 | (MP_CODE_NATIVE_VIPER - MP_CODE_BYTECODE))
# 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 07:22:47 -04:00
# MPY: bss and/or rodata
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 07:22:47 -04:00
bss_const_table_idx = 2
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 07:22:47 -04:00
if len(env.full_rodata):
out.write_bytes(env.full_rodata)
# MPY: relocation information
prev_kind = None
for base, addr, kind in env.mpy_relocs:
if isinstance(kind, str) and kind.startswith(".text"):
kind = 0
elif kind in (".rodata", ".data.rel.ro"):
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
elif kind == "mp_native_qstr_table":
kind = 6
elif kind == "mp_native_obj_table":
kind = 7
elif kind == "mp_fun_table":
kind = 8
else:
kind = 9 + kind
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()
################################################################################
# main
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)):
print("#define %s (mp_native_qstr_table[%d])" % (q, i + 1), file=f)
for i, q in enumerate(sorted(qstr_objs)):
print(
"#define MP_OBJ_NEW_QSTR_%s ((mp_obj_t)mp_native_obj_table[%d])" % (q, i),
file=f,
)
print("extern const uint16_t mp_native_qstr_table[];", file=f)
print("extern const mp_uint_t mp_native_obj_table[];", file=f)
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)
def main():
import argparse
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)
if __name__ == "__main__":
main()