Skip to content
Switch branches/tags
Go to file
Cannot retrieve contributors at this time
import warnings
import functools
import locale
import weakref
import ctypes
import html
import textwrap
import llvmlite.llvmpy.core as lc
import llvmlite.llvmpy.passes as lp
import llvmlite.binding as ll
import as llvmir
from abc import abstractmethod, ABCMeta
from numba.core import utils, config, cgutils
from numba.core.runtime.nrtopt import remove_redundant_nrt_refct
from numba.core.runtime import rtsys
from numba.core.compiler_lock import require_global_compiler_lock
from numba.core.errors import NumbaInvalidConfigWarning
from numba.misc.inspection import disassemble_elf_to_cfg
from numba.misc.llvm_pass_timings import PassTimingsCollection
_x86arch = frozenset(['x86', 'i386', 'i486', 'i586', 'i686', 'i786',
'i886', 'i986'])
def _is_x86(triple):
arch = triple.split('-')[0]
return arch in _x86arch
def _parse_refprune_flags():
"""Parse refprune flags from the `config`.
Invalid values are ignored an warn via a `NumbaInvalidConfigWarning`
flags : llvmlite.binding.RefPruneSubpasses
flags = config.LLVM_REFPRUNE_FLAGS.split(',')
if not flags:
return 0
val = 0
for item in flags:
item = item.strip()
val |= getattr(ll.RefPruneSubpasses, item.upper())
except AttributeError:
warnings.warn(f"invalid refprune flags {item!r}",
return val
def dump(header, body, lang):
import pygments
except ImportError:
msg = "Please install pygments to see highlighted dumps"
raise ValueError(msg)
from pygments import highlight
from pygments.lexers import GasLexer as gas_lexer
from pygments.lexers import LlvmLexer as llvm_lexer
from pygments.formatters import Terminal256Formatter
from numba.misc.dump_style import by_colorscheme
lexer_map = {'llvm': llvm_lexer, 'asm': gas_lexer}
lexer = lexer_map[lang]
def printer(arg):
print(highlight(arg, lexer(),
printer = print
print('=' * 80)
print(, '-'))
print('=' * 80)
class _CFG(object):
Wraps the CFG graph for different display method.
Instance of the class can be stringified (``__repr__`` is defined) to get
the graph in DOT format. The ``.display()`` method plots the graph in
PDF. If in IPython notebook, the returned image can be inlined.
def __init__(self, cres, name, py_func, **kwargs):
self.cres = cres = name
self.py_func = py_func
fn = cres.get_function(name) = ll.get_function_cfg(fn)
self.kwargs = kwargs
def pretty_printer(self, filename=None, view=None, render_format=None,
interleave=False, strip_ir=False, show_key=True,
"Pretty" prints the DOT graph of the CFG.
For explanation of the parameters see the docstring for
import graphviz as gv
import re
import json
import inspect
from llvmlite import binding as ll
from numba.typed import List
from types import SimpleNamespace
from collections import defaultdict
_default = False
_highlight = SimpleNamespace(incref=_default,
_interleave = SimpleNamespace(python=_default, lineinfo=_default)
def parse_config(_config, kwarg):
""" Parses the kwarg into a consistent format for use in configuring
the Digraph rendering. _config is the configuration instance to
update, kwarg is the kwarg on which to base the updates.
if isinstance(kwarg, bool):
for attr in _config.__dict__:
setattr(_config, attr, kwarg)
elif isinstance(kwarg, dict):
for k, v in kwarg.items():
if k not in _config.__dict__:
raise ValueError("Unexpected key in kwarg: %s" % k)
if isinstance(v, bool):
setattr(_config, k, v)
msg = "Unexpected value for key: %s, got:%s"
raise ValueError(msg % (k, v))
elif isinstance(kwarg, set):
for item in kwarg:
if item not in _config.__dict__:
raise ValueError("Unexpected key in kwarg: %s" % item)
setattr(_config, item, True)
msg = "Unhandled configuration type for kwarg %s"
raise ValueError(msg % type(kwarg))
parse_config(_highlight, highlight)
parse_config(_interleave, interleave)
# This is the colour scheme. The graphviz HTML label renderer only takes
# names for colours:
cs = defaultdict(lambda: 'white') # default bg colour is white
cs['marker'] = 'orange'
cs['python'] = 'yellow'
cs['truebr'] = 'green'
cs['falsebr'] = 'red'
cs['incref'] = 'cyan'
cs['decref'] = 'turquoise'
cs['raise'] = 'lightpink'
cs['meminfo'] = 'lightseagreen'
cs['return'] = 'purple'
cs['llvm_intrin_calls'] = 'rosybrown'
cs['function_calls'] = 'tomato'
# Get the raw dot format information from LLVM and the LLVM IR
fn = self.cres.get_function(
#raw_dot = ll.get_function_cfg(fn).replace('\\l...', '')
llvm_str = self.cres.get_llvm_str()
def get_metadata(llvm_str):
""" Gets the metadata entries from the LLVM IR, these look something
like '!123 = INFORMATION'. Returns a map of metadata key to metadata
value, i.e. from the example {'!123': INFORMATION}"""
md = {}
metadata_entry = re.compile(r'(^[!][0-9]+)(\s+=\s+.*)')
for x in llvm_str.splitlines():
match = metadata_entry.match(x)
if match is not None:
g = match.groups()
if g is not None:
assert len(g) == 2
md[g[0]] = g[1]
return md
md = get_metadata(llvm_str)
# setup digraph with initial properties
def init_digraph(name, fname, fontsize):
# name and fname are arbitrary graph and file names, they appear in
# some rendering formats, the fontsize determines the output
# fontsize.
# truncate massive mangled names as file names as it causes OSError
# when trying to render to pdf
cmax = 200
if len(fname) > cmax:
wstr = (f'CFG output filname "{fname}" exceeds maximum '
f'supported length, it will be truncated.')
warnings.warn(wstr, NumbaInvalidConfigWarning)
fname = fname[:cmax]
f = gv.Digraph(name, filename=fname)
f.attr('node', shape='none', fontsize='%s' % str(fontsize))
return f
f = init_digraph(,, fontsize)
# A lot of regex is needed to parse the raw dot output. This output
# contains a mix of LLVM IR in the labels, and also DOT markup.
# DOT syntax, matches a "port" (where the tail of an edge starts)
port_match = re.compile('.*{(.*)}.*')
# DOT syntax, matches the "port" value from a found "port_match"
port_jmp_match = re.compile('.*<(.*)>(.*)')
# LLVM syntax, matches a LLVM debug marker
metadata_marker = re.compile(r'.*!dbg\s+(![0-9]+).*')
# LLVM syntax, matches a location entry
location_expr = (r'.*!DILocation\(line:\s+([0-9]+),'
location_entry = re.compile(location_expr)
# LLVM syntax, matches LLVMs internal debug value calls
dbg_value = re.compile(r'.*call void @llvm.dbg.value.*')
# LLVM syntax, matches tokens for highlighting
nrt_incref = re.compile(r"@NRT_incref\b")
nrt_decref = re.compile(r"@NRT_decref\b")
nrt_meminfo = re.compile("@NRT_MemInfo")
ll_intrin_calls = re.compile(r".*call.*@llvm\..*")
ll_function_call = re.compile(r".*call.*@.*")
ll_raise = re.compile(r"ret i32.*\!ret_is_raise.*")
ll_return = re.compile("ret i32 [^1],?.*")
# wrapper function for line wrapping LLVM lines
def wrap(s):
return textwrap.wrap(s, width=120, subsequent_indent='... ')
# function to fix (sometimes escaped for DOT!) LLVM IR etc that needs to
# be HTML escaped
def clean(s):
# Grab first 300 chars only, 1. this should be enough to identify
# the token and it keeps names short. 2. graphviz/dot has a maximum
# buffer size near 585?!, with additional transforms it's hard to
# know if this would be exceeded. 3. hash of the token string is
# written into the rendering to permit exact identification against
# e.g. LLVM IR dump if necessary.
n = 300
if len(s) > n:
hs = str(hash(s))
s = '{}...<hash={}>'.format(s[:n], hs)
s = html.escape(s) # deals with &, < and >
s = s.replace('\\{', "&#123;")
s = s.replace('\\}', "&#125;")
s = s.replace('\\', "&#92;")
s = s.replace('%', "&#37;")
s = s.replace('!', "&#33;")
return s
# These hold the node and edge ids from the raw dot information. They
# are used later to wire up a new DiGraph that has the same structure
# as the raw dot but with new nodes.
node_ids = {}
edge_ids = {}
# Python source lines, used if python source interleave is requested
if _interleave.python:
src_code, firstlineno = inspect.getsourcelines(self.py_func)
# This is the dot info from LLVM, it's in DOT form and has continuation
# lines, strip them and then re-parse into `dot_json` form for use in
# producing a formatted output.
raw_dot = ll.get_function_cfg(fn).replace('\\l...', '')
json_bytes = gv.Source(raw_dot).pipe(format='dot_json')
jzon = json.loads(json_bytes.decode('utf-8'))
idc = 0
# Walk the "objects" (nodes) in the DOT output
for obj in jzon['objects']:
# These are used to keep tabs on the current line and column numbers
# as per the markers. They are tracked so as to make sure a marker
# is only emitted if there's a change in the marker.
cur_line, cur_col = -1, -1
label = obj['label']
name = obj['name']
gvid = obj['_gvid']
node_ids[gvid] = name
# Label is DOT format, it needs the head and tail removing and then
# splitting for walking.
label = label[1:-1]
lines = label.split('\\l')
# Holds the new lines
new_lines = []
# Aim is to produce an HTML table a bit like this:
# |------------|
# | HEADER | <-- this is the block header
# |------------|
# | LLVM SRC | <--
# | Marker? | < this is the label/block body
# | Python src?| <--
# |------------|
# | T | F | <-- this is the "ports", also determines col_span
# --------------
# This is HTML syntax, its the column span. If there's a switch or a
# branch at the bottom of the node this is rendered as multiple
# columns in a table. First job is to go and render that and work
# out how many columns are needed as that dictates how many columns
# the rest of the source lines must span. In DOT syntax the places
# that edges join nodes are referred to as "ports". Syntax in DOT
# is like `node:port`.
col_span = 1
# First see if there is a port entry for this node
port_line = ''
matched = port_match.match(lines[-1])
sliced_lines = lines
if matched is not None:
# There is a port
ports = matched.groups()[0]
ports_tokens = ports.split('|')
col_span = len(ports_tokens)
# Generate HTML table data cells, one for each port. If the
# ports correspond to a branch then they can optionally
# highlighted based on T/F.
tdfmt = ('<td BGCOLOR="{}" BORDER="1" ALIGN="center" '
tbl_data = []
if _highlight.branches:
colors = {'T': cs['truebr'], 'F': cs['falsebr']}
colors = {}
for tok in ports_tokens:
target, value = port_jmp_match.match(tok).groups()
color = colors.get(value, 'white')
tbl_data.append(tdfmt.format(color, target, value))
port_line = ''.join(tbl_data)
# Drop the last line from the rest of the parse as it's the port
# and just been dealt with.
sliced_lines = lines[:-1]
# loop peel the block header, it needs a HTML border
fmtheader = ('<tr><td BGCOLOR="{}" BORDER="1" ALIGN="left" '
new_lines.append(fmtheader.format(cs['default'], col_span,
# process rest of block creating the table row at a time.
fmt = ('<tr><td BGCOLOR="{}" BORDER="0" ALIGN="left" '
def metadata_interleave(l, new_lines):
Search line `l` for metadata associated with python or line info
and inject it into `new_lines` if requested.
matched = metadata_marker.match(l)
if matched is not None:
# there's a metadata marker
g = matched.groups()
if g is not None:
assert len(g) == 1, g
marker = g[0]
debug_data = md.get(marker, None)
if debug_data is not None:
# and the metadata marker has a corresponding piece
# of metadata
ld = location_entry.match(debug_data)
if ld is not None:
# and the metadata is line info... proceed
assert len(ld.groups()) == 2, ld
line, col = ld.groups()
# only emit a new marker if the line number in
# the metadata is "new".
if line != cur_line or col != cur_col:
if _interleave.lineinfo:
mfmt = 'Marker %s, Line %s, column %s'
mark_line = mfmt % (marker, line, col)
ln = fmt.format(cs['marker'], col_span,
if _interleave.python:
# +1 for decorator, this probably needs
# the same thing doing as for the
# error messages where the decorator
# is scanned for, its not always +1!
lidx = int(line) - (firstlineno + 1)
source_line = src_code[lidx + 1]
ln = fmt.format(cs['python'], col_span,
return line, col
for l in sliced_lines[1:]:
# Drop LLVM debug call entries
if dbg_value.match(l):
# if requested generate interleaving of markers or python from
# metadata
if _interleave.lineinfo or _interleave.python:
updated_lineinfo = metadata_interleave(l, new_lines)
if updated_lineinfo is not None:
cur_line, cur_col = updated_lineinfo
# Highlight other LLVM features if requested, HTML BGCOLOR
# property is set by this.
if _highlight.incref and
colour = cs['incref']
elif _highlight.decref and
colour = cs['decref']
elif _highlight.meminfo and
colour = cs['meminfo']
elif _highlight.raises and
# search for raise as its more specific than exit
colour = cs['raise']
elif _highlight.returns and
colour = cs['return']
elif _highlight.llvm_intrin_calls and
colour = cs['llvm_intrin_calls']
elif _highlight.function_calls and
colour = cs['function_calls']
colour = cs['default']
# Use the default coloring as a flag to force printing if a
# special token print was requested AND LLVM ir stripping is
# required
if colour is not cs['default'] or not strip_ir:
for x in wrap(clean(l)):
new_lines.append(fmt.format(colour, col_span, x))
# add in the port line at the end of the block if it was present
# (this was built right at the top of the parse)
if port_line:
# If there was data, create a table, else don't!
dat = ''.join(new_lines)
if dat:
tab = (('<table id="%s" BORDER="1" CELLBORDER="0" '
'CELLPADDING="0" CELLSPACING="0">%s</table>') % (idc,
label = '<{}>'.format(tab)
label = ''
# finally, add a replacement node for the original with a new marked
# up label.
f.node(name, label=label)
# Parse the edge data
if 'edges' in jzon: # might be a single block, no edges
for edge in jzon['edges']:
gvid = edge['_gvid']
tp = edge.get('tailport', None)
edge_ids[gvid] = (edge['head'], edge['tail'], tp)
# Write in the edge wiring with respect to the new nodes:ports.
for gvid, edge in edge_ids.items():
tail = node_ids[edge[1]]
head = node_ids[edge[0]]
port = edge[2]
if port is not None:
tail += ':%s' % port
f.edge(tail, head)
# Add a key to the graph if requested.
if show_key:
key_tab = []
for k, v in cs.items():
key_tab.append(('<tr><td BGCOLOR="{}" BORDER="0" ALIGN="center"'
'>{}</td></tr>').format(v, k))
# The first < and last > are DOT syntax, rest is DOT HTML.
f.node("Key", label=('<<table BORDER="1" CELLBORDER="1" '
# Render if required
if filename is not None or view is not None:
f.render(filename=filename, view=view, format=render_format)
# Else pipe out a SVG
return f.pipe(format='svg')
def display(self, filename=None, format='pdf', view=False):
Plot the CFG. In IPython notebook, the return image object can be
The *filename* option can be set to a specific path for the rendered
output to write to. If *view* option is True, the plot is opened by
the system default application for the image format (PDF). *format* can
be any valid format string accepted by graphviz, default is 'pdf'.
rawbyt = self.pretty_printer(filename=filename, view=view,
render_format=format, **self.kwargs)
return rawbyt.decode('utf-8')
def _repr_svg_(self):
return self.pretty_printer(**self.kwargs).decode('utf-8')
def __repr__(self):
class CodeLibrary(metaclass=ABCMeta):
An interface for bundling LLVM code together and compiling it.
It is tied to a *codegen* instance (e.g. JITCPUCodegen) that will
determine how the LLVM code is transformed and linked together.
_finalized = False
_object_caching_enabled = False
_disable_inspection = False
def __init__(self, codegen: "CPUCodegen", name: str):
self._codegen = codegen
self._name = name
ptc_name = f"{self.__class__.__name__}({self._name!r})"
self._recorded_timings = PassTimingsCollection(ptc_name)
# Track names of the dynamic globals
self._dynamic_globals = []
def has_dynamic_globals(self):
return len(self._dynamic_globals) > 0
def recorded_timings(self):
return self._recorded_timings
def codegen(self):
The codegen object owning this library.
return self._codegen
def name(self):
return self._name
def __repr__(self):
return "<Library %r at 0x%x>" % (, id(self))
def _raise_if_finalized(self):
if self._finalized:
raise RuntimeError("operation impossible on finalized object %r"
% (self,))
def _ensure_finalized(self):
if not self._finalized:
def create_ir_module(self, name):
Create an LLVM IR module for use by this library.
ir_module = self._codegen._create_empty_module(name)
return ir_module
def add_linking_library(self, library):
Add a library for linking into this library, without losing
the original library.
def add_ir_module(self, ir_module):
Add an LLVM IR module's contents to this library.
def finalize(self):
Finalize the library. After this call, nothing can be added anymore.
Finalization involves various stages of code optimization and
def get_function(self, name):
Return the function named ``name``.
def get_llvm_str(self):
Get the human-readable form of the LLVM module.
def get_asm_str(self):
Get the human-readable assembly.
# Object cache hooks and serialization
def enable_object_caching(self):
self._object_caching_enabled = True
self._compiled_object = None
self._compiled = False
def _get_compiled_object(self):
if not self._object_caching_enabled:
raise ValueError("object caching not enabled in %s" % (self,))
if self._compiled_object is None:
raise RuntimeError("no compiled object yet for %s" % (self,))
return self._compiled_object
def _set_compiled_object(self, value):
if not self._object_caching_enabled:
raise ValueError("object caching not enabled in %s" % (self,))
if self._compiled:
raise ValueError("library already compiled: %s" % (self,))
self._compiled_object = value
self._disable_inspection = True
class CPUCodeLibrary(CodeLibrary):
def __init__(self, codegen, name):
super().__init__(codegen, name)
self._linking_libraries = [] # maintain insertion order
self._final_module = ll.parse_assembly(
str(self._codegen._create_empty_module( = cgutils.normalize_ir_text(
self._shared_module = None
def _optimize_functions(self, ll_module):
Internal: run function-level optimizations inside *ll_module*.
# Enforce data layout to enable layout-specific optimizations
ll_module.data_layout = self._codegen._data_layout
with self._codegen._function_pass_manager(ll_module) as fpm:
# Run function-level optimizations to reduce memory usage and improve
# module-level optimization.
for func in ll_module.functions:
k = f"Function passes on {!r}"
with self._recorded_timings.record(k):
def _optimize_final_module(self):
Internal: optimize this library's final module.
cheap_name = "Module passes (cheap optimization for refprune)"
with self._recorded_timings.record(cheap_name):
# A cheaper optimisation pass is run first to try and get as many
# refops into the same function as possible via inlining
# Refop pruning is then run on the heavily inlined function
if not config.LLVM_REFPRUNE_PASS:
self._final_module = remove_redundant_nrt_refct(self._final_module)
full_name = "Module passes (full optimization)"
with self._recorded_timings.record(full_name):
# The full optimisation suite is then run on the refop pruned IR
def _get_module_for_linking(self):
Internal: get a LLVM module suitable for linking multiple times
into another library. Exported functions are made "linkonce_odr"
to allow for multiple definitions, inlining, and removal of
unused exports.
See discussion in
if self._shared_module is not None:
return self._shared_module
mod = self._final_module
to_fix = []
nfuncs = 0
for fn in mod.functions:
nfuncs += 1
if not fn.is_declaration and fn.linkage == ll.Linkage.external:
if nfuncs == 0:
# This is an issue which can occur if loading a module
# from an object file and trying to link with it, so detect it
# here to make debugging easier.
raise RuntimeError("library unfit for linking: "
"no available functions in %s"
% (self,))
if to_fix:
mod = mod.clone()
for name in to_fix:
# NOTE: this will mark the symbol WEAK if serialized
# to an ELF file
mod.get_function(name).linkage = 'linkonce_odr'
self._shared_module = mod
return mod
def add_linking_library(self, library):
def add_ir_module(self, ir_module):
assert isinstance(ir_module, llvmir.Module)
ir = cgutils.normalize_ir_text(str(ir_module))
ll_module = ll.parse_assembly(ir) =
def add_llvm_module(self, ll_module):
# TODO: we shouldn't need to recreate the LLVM module object
if not config.LLVM_REFPRUNE_PASS:
ll_module = remove_redundant_nrt_refct(ll_module)
def finalize(self):
# Report any LLVM-related problems to the user
if config.DUMP_FUNC_OPT:
self.get_llvm_str(), 'llvm')
# Link libraries for shared code
seen = set()
for library in self._linking_libraries:
if library not in seen:
library._get_module_for_linking(), preserve=True,
# Optimize the module after all dependences are linked in above,
# to allow for inlining.
def _finalize_dynamic_globals(self):
# Scan for dynamic globals
for gv in self._final_module.global_variables:
def _verify_declare_only_symbols(self):
# Verify that no declare-only function compiled by numba.
for fn in self._final_module.functions:
# We will only check for symbol name starting with '_ZN5numba'
if fn.is_declaration and'_ZN5numba'):
msg = 'Symbol {} not linked properly'
raise AssertionError(msg.format(
def _finalize_final_module(self):
Make the underlying LLVM module ready to use.
# Remember this on the module, for the object cache hooks
self._final_module.__library = weakref.proxy(self)
# It seems add_module() must be done only here and not before
# linking in other modules, otherwise get_pointer_to_function()
# could fail.
cleanup = self._codegen._add_module(self._final_module)
if cleanup:
weakref.finalize(self, cleanup)
self._finalized = True
dump("OPTIMIZED DUMP %s" %, self.get_llvm_str(), 'llvm')
if config.DUMP_ASSEMBLY:
dump("ASSEMBLY %s" %, self.get_asm_str(), 'asm')
def get_defined_functions(self):
Get all functions defined in the library. The library must have
been finalized.
mod = self._final_module
for fn in mod.functions:
if not fn.is_declaration:
yield fn
def get_function(self, name):
return self._final_module.get_function(name)
def _sentry_cache_disable_inspection(self):
if self._disable_inspection:
warnings.warn('Inspection disabled for cached code. '
'Invalid result is returned.')
def get_llvm_str(self):
return str(self._final_module)
def get_asm_str(self):
return str(self._codegen._tm.emit_assembly(self._final_module))
def get_function_cfg(self, name, py_func=None, **kwargs):
Get control-flow graph of the LLVM function
return _CFG(self, name, py_func, **kwargs)
def get_disasm_cfg(self, mangled_name):
Get the CFG of the disassembly of the ELF object at symbol mangled_name.
Requires python package: r2pipe
Requires radare2 binary on $PATH.
Notebook rendering requires python package: graphviz
Optionally requires a compiler toolchain (via pycc) to link the ELF to
get better disassembly results.
elf = self._get_compiled_object()
return disassemble_elf_to_cfg(elf, mangled_name)
def _dump_elf(cls, buf):
Dump the symbol table of an ELF file.
Needs pyelftools (
from elftools.elf.elffile import ELFFile
from elftools.elf import descriptions
from io import BytesIO
f = ELFFile(BytesIO(buf))
print("ELF file:")
for sec in f.iter_sections():
if sec['sh_type'] == 'SHT_SYMTAB':
symbols = sorted(sec.iter_symbols(), key=lambda sym:
print(" symbols:")
for sym in symbols:
if not
print(" - %r: size=%d, value=0x%x, type=%s, bind=%s"
% (,
def _object_compiled_hook(cls, ll_module, buf):
`ll_module` was compiled into object code `buf`.
self = ll_module.__library
except AttributeError:
if self._object_caching_enabled:
self._compiled = True
self._compiled_object = buf
def _object_getbuffer_hook(cls, ll_module):
Return a cached object code for `ll_module`.
self = ll_module.__library
except AttributeError:
if self._object_caching_enabled and self._compiled_object:
buf = self._compiled_object
self._compiled_object = None
return buf
def serialize_using_bitcode(self):
Serialize this library using its bitcode as the cached representation.
return (, 'bitcode', self._final_module.as_bitcode())
def serialize_using_object_code(self):
Serialize this library using its object code as the cached
representation. We also include its bitcode for further inlining
with other libraries.
data = (self._get_compiled_object(),
return (, 'object', data)
def _unserialize(cls, codegen, state):
name, kind, data = state
self = codegen.create_library(name)
assert isinstance(self, cls)
if kind == 'bitcode':
# No need to re-run optimizations, just make the module ready
self._final_module = ll.parse_bitcode(data)
return self
elif kind == 'object':
object_code, shared_bitcode = data
self._shared_module = ll.parse_bitcode(shared_bitcode)
# Load symbols from cache
return self
raise ValueError("unsupported serialization kind %r" % (kind,))
class AOTCodeLibrary(CPUCodeLibrary):
def emit_native_object(self):
Return this library as a native object (a bytestring) -- for example
ELF under Linux.
This function implicitly calls .finalize().
return self._codegen._tm.emit_object(self._final_module)
def emit_bitcode(self):
Return this library as LLVM bitcode (a bytestring).
This function implicitly calls .finalize().
return self._final_module.as_bitcode()
def _finalize_specific(self):
class JITCodeLibrary(CPUCodeLibrary):
def get_pointer_to_function(self, name):
Generate native code for function named *name* and return a pointer
to the start of the function (as an integer).
This function implicitly calls .finalize().
pointer : int
- zero (null) if no symbol of *name* is defined by this code
- non-zero if the symbol is defined.
ee = self._codegen._engine
if not ee.is_symbol_defined(name):
return 0
return self._codegen._engine.get_function_address(name)
def _finalize_specific(self):
with self._recorded_timings.record("Finalize object"):
class RuntimeLinker(object):
For tracking unresolved symbols generated at runtime due to recursion.
PREFIX = '.numba.unresolved$'
def __init__(self):
self._unresolved = utils.UniqueDict()
self._defined = set()
self._resolved = []
def scan_unresolved_symbols(self, module, engine):
Scan and track all unresolved external symbols in the module and
allocate memory for it.
prefix = self.PREFIX
for gv in module.global_variables:
sym =[len(prefix):]
# Avoid remapping to existing GV
if engine.is_symbol_defined(
# Allocate a memory space for the pointer
abortfn = rtsys.library.get_pointer_to_function("nrt_unresolved_abort")
ptr = ctypes.c_void_p(abortfn)
engine.add_global_mapping(gv, ctypes.addressof(ptr))
self._unresolved[sym] = ptr
def scan_defined_symbols(self, module):
Scan and track all defined symbols.
for fn in module.functions:
if not fn.is_declaration:
def resolve(self, engine):
Fix unresolved symbols if they are defined.
# An iterator to get all unresolved but available symbols
pending = [name for name in self._unresolved if name in self._defined]
# Resolve pending symbols
for name in pending:
# Get runtime address
fnptr = engine.get_function_address(name)
# Fix all usage
ptr = self._unresolved[name]
ptr.value = fnptr
self._resolved.append((name, ptr)) # keep ptr alive
# Delete resolved
del self._unresolved[name]
def _proxy(old):
def wrapper(self, *args, **kwargs):
return old(self._ee, *args, **kwargs)
return wrapper
class JitEngine(object):
"""Wraps an ExecutionEngine to provide custom symbol tracking.
Since the symbol tracking is incomplete (doesn't consider
loaded code object), we are not putting it in llvmlite.
def __init__(self, ee):
self._ee = ee
# Track symbol defined via codegen'd Module
# but not any cached object.
# NOTE: `llvm::ExecutionEngine` will catch duplicated symbols and
# we are not going to protect against that. A proper duplicated
# symbol detection will need a more logic to check for the linkage
# (e.g. like `weak` linkage symbol can override). This
# `_defined_symbols` set will be just enough to tell if a symbol
# exists and will not cause the `EE` symbol lookup to `exit(1)`
# when symbol-not-found.
self._defined_symbols = set()
def is_symbol_defined(self, name):
"""Is the symbol defined in this session?
return name in self._defined_symbols
def _load_defined_symbols(self, mod):
"""Extract symbols from the module
for gsets in (mod.functions, mod.global_variables):
self._defined_symbols |= { for gv in gsets
if not gv.is_declaration}
def add_module(self, module):
"""Override ExecutionEngine.add_module
to keep info about defined symbols.
return self._ee.add_module(module)
def add_global_mapping(self, gv, addr):
"""Override ExecutionEngine.add_global_mapping
to keep info about defined symbols.
return self._ee.add_global_mapping(gv, addr)
# The remaining methods are re-export of the ExecutionEngine APIs
set_object_cache = _proxy(ll.ExecutionEngine.set_object_cache)
finalize_object = _proxy(ll.ExecutionEngine.finalize_object)
get_function_address = _proxy(ll.ExecutionEngine.get_function_address)
get_global_value_address = _proxy(
class Codegen(metaclass=ABCMeta):
Base Codegen class. It is expected that subclasses set the class attribute
``_library_class``, indicating the CodeLibrary class for the target.
Subclasses should also initialize:
``self._data_layout``: the data layout for the target.
``self._target_data``: the binding layer ``TargetData`` for the target.
def _create_empty_module(self, name):
Create a new empty module suitable for the target.
def _add_module(self, module):
Add a module to the execution engine. Ownership of the module is
transferred to the engine.
def target_data(self):
The LLVM "target data" object for this codegen instance.
return self._target_data
def create_library(self, name, **kwargs):
Create a :class:`CodeLibrary` object for use with this codegen
return self._library_class(self, name, **kwargs)
def unserialize_library(self, serialized):
return self._library_class._unserialize(self, serialized)
class CPUCodegen(Codegen):
def __init__(self, module_name):
self._data_layout = None
self._llvm_module = ll.parse_assembly(
str(self._create_empty_module(module_name))) = "global_codegen_module"
self._rtlinker = RuntimeLinker()
def _init(self, llvm_module):
assert list(llvm_module.global_variables) == [], "Module isn't empty"
target = ll.Target.from_triple(ll.get_process_triple())
tm_options = dict(opt=config.OPT)
self._tm_features = self._customize_tm_features()
tm = target.create_target_machine(**tm_options)
engine = ll.create_mcjit_compiler(llvm_module, tm)
self._tm = tm
self._engine = JitEngine(engine)
self._target_data = engine.target_data
self._data_layout = str(self._target_data)
self._mpm_cheap = self._module_pass_manager(loop_vectorize=False,
self._mpm_full = self._module_pass_manager()
def _create_empty_module(self, name):
ir_module = llvmir.Module(cgutils.normalize_ir_text(name))
ir_module.triple = ll.get_process_triple()
if self._data_layout:
ir_module.data_layout = self._data_layout
return ir_module
def _module_pass_manager(self, **kwargs):
pm = ll.create_module_pass_manager()
cost = kwargs.pop("cost", None)
with self._pass_manager_builder(**kwargs) as pmb:
# If config.OPT==0 do not include these extra passes to help with
# vectorization.
if cost is not None and cost == "cheap" and config.OPT != 0:
# This knocks loops into rotated form early to reduce the likelihood
# of vectorization failing due to unknown PHI nodes.
# LLVM 11 added LFTR to the IV Simplification pass, this interacted
# badly with the existing use of the InstructionCombiner here and
# ended up with PHI nodes that prevented vectorization from
# working. The desired vectorization effects can be achieved
# with this in LLVM 11 (and also < 11) but at a potentially
# slightly higher cost:
return pm
def _function_pass_manager(self, llvm_module, **kwargs):
pm = ll.create_function_pass_manager(llvm_module)
with self._pass_manager_builder(**kwargs) as pmb:
return pm
def _pass_manager_builder(self, **kwargs):
Create a PassManagerBuilder.
Note: a PassManagerBuilder seems good only for one use, so you
should call this method each time you want to populate a module
or function pass manager. Otherwise some optimizations will be
opt_level = kwargs.pop('opt', config.OPT)
loop_vectorize = kwargs.pop('loop_vectorize', config.LOOP_VECTORIZE)
slp_vectorize = kwargs.pop('slp_vectorize', config.SLP_VECTORIZE)
pmb = lp.create_pass_manager_builder(opt=opt_level,
return pmb
def _check_llvm_bugs(self):
Guard against some well-known LLVM bug(s).
# Check the locale bug at
# Note we can't cache the result as locale settings can change
# across a process's lifetime. Also, for this same reason,
# the check here is a mere heuristic (there may be a race condition
# between now and actually compiling IR).
ir = """
define double @func()
ret double 1.23e+01
mod = ll.parse_assembly(ir)
ir_out = str(mod)
if "12.3" in ir_out or "1.23" in ir_out:
# Everything ok
if "1.0" in ir_out:
loc = locale.getlocale()
raise RuntimeError(
"LLVM will produce incorrect floating-point code "
"in the current locale %s.\nPlease read "
" "
"for more information."
% (loc,))
raise AssertionError("Unexpected IR:\n%s\n" % (ir_out,))
def magic_tuple(self):
Return a tuple unambiguously describing the codegen behaviour.
return (self._llvm_module.triple, self._get_host_cpu_name(),
def _scan_and_fix_unresolved_refs(self, module):
self._rtlinker.scan_unresolved_symbols(module, self._engine)
def insert_unresolved_ref(self, builder, fnty, name):
voidptr = llvmir.IntType(8).as_pointer()
ptrname = self._rtlinker.PREFIX + name
llvm_mod = builder.module
fnptr = llvm_mod.get_global(ptrname)
except KeyError:
# Not defined?
fnptr = llvmir.GlobalVariable(llvm_mod, voidptr, name=ptrname)
fnptr.linkage = 'external'
return builder.bitcast(builder.load(fnptr), fnty.as_pointer())
def _get_host_cpu_name(self):
return (ll.get_host_cpu_name()
if config.CPU_NAME is None
else config.CPU_NAME)
def _get_host_cpu_features(self):
if config.CPU_FEATURES is not None:
return config.CPU_FEATURES
return get_host_cpu_features()
class AOTCPUCodegen(CPUCodegen):
A codegen implementation suitable for Ahead-Of-Time compilation
(e.g. generation of object files).
_library_class = AOTCodeLibrary
def __init__(self, module_name, cpu_name=None):
# By default, use generic cpu model for the arch
self._cpu_name = cpu_name or ''
CPUCodegen.__init__(self, module_name)
def _customize_tm_options(self, options):
cpu_name = self._cpu_name
if cpu_name == 'host':
cpu_name = self._get_host_cpu_name()
options['cpu'] = cpu_name
options['reloc'] = 'pic'
options['codemodel'] = 'default'
options['features'] = self._tm_features
def _customize_tm_features(self):
# ISA features are selected according to the requested CPU model
# in _customize_tm_options()
return ''
def _add_module(self, module):
class JITCPUCodegen(CPUCodegen):
A codegen implementation suitable for Just-In-Time compilation.
_library_class = JITCodeLibrary
def _customize_tm_options(self, options):
# As long as we don't want to ship the code to another machine,
# we can specialize for this CPU.
options['cpu'] = self._get_host_cpu_name()
# LLVM 7 change: #
# JIT needs static relocation on x86*
# native target is already initialized from base class __init__
arch = ll.Target.from_default_triple().name
if arch.startswith('x86'): # one of x86 or x86_64
reloc_model = 'static'
elif arch.startswith('ppc'):
reloc_model = 'pic'
reloc_model = 'default'
options['reloc'] = reloc_model
options['codemodel'] = 'jitdefault'
# Set feature attributes (such as ISA extensions)
# This overrides default feature selection by CPU model above
options['features'] = self._tm_features
# Deal with optional argument to ll.Target.create_target_machine
sig = utils.pysignature(ll.Target.create_target_machine)
if 'jit' in sig.parameters:
# Mark that this is making a JIT engine
options['jit'] = True
def _customize_tm_features(self):
# For JIT target, we will use LLVM to get the feature map
return self._get_host_cpu_features()
def _add_module(self, module):
# XXX: disabling remove module due to MCJIT engine leakage in
# removeModule. The removeModule causes consistent access
# violation with certain test combinations.
# # Early bind the engine method to avoid keeping a reference to self.
# return functools.partial(self._engine.remove_module, module)
def set_env(self, env_name, env):
"""Set the environment address.
Update the GlobalVariable named *env_name* to the address of *env*.
gvaddr = self._engine.get_global_value_address(env_name)
envptr = (ctypes.c_void_p * 1).from_address(gvaddr)
envptr[0] = ctypes.c_void_p(id(env))
def initialize_llvm():
"""Safe to use multiple times.
def get_host_cpu_features():
"""Get host CPU features using LLVM.
The features may be modified due to user setting.
See numba.config.ENABLE_AVX.
features = ll.get_host_cpu_features()
except RuntimeError:
return ''
if not config.ENABLE_AVX:
# Disable all features with name starting with 'avx'
for k in features:
if k.startswith('avx'):
features[k] = False
# Set feature attributes
return features.flatten()