artiq/artiq/compiler/targets.py
2021-01-11 19:43:12 +08:00

284 lines
10 KiB
Python

import os, sys, tempfile, subprocess, io
from artiq.compiler import types, ir
from llvmlite_artiq import ir as ll, binding as llvm
llvm.initialize()
llvm.initialize_all_targets()
llvm.initialize_all_asmprinters()
class RunTool:
def __init__(self, pattern, **tempdata):
self._pattern = pattern
self._tempdata = tempdata
self._tempnames = {}
self._tempfiles = {}
def __enter__(self):
for key, data in self._tempdata.items():
if data is None:
fd, filename = tempfile.mkstemp()
os.close(fd)
self._tempnames[key] = filename
else:
with tempfile.NamedTemporaryFile(delete=False) as f:
f.write(data)
self._tempnames[key] = f.name
cmdline = []
for argument in self._pattern:
cmdline.append(argument.format(**self._tempnames))
process = subprocess.Popen(cmdline, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
universal_newlines=True)
stdout, stderr = process.communicate()
if process.returncode != 0:
raise Exception("{} invocation failed: {}".
format(cmdline[0], stderr))
self._tempfiles["__stdout__"] = io.StringIO(stdout)
for key in self._tempdata:
if self._tempdata[key] is None:
self._tempfiles[key] = open(self._tempnames[key], "rb")
return self._tempfiles
def __exit__(self, exc_typ, exc_value, exc_trace):
for file in self._tempfiles.values():
file.close()
for filename in self._tempnames.values():
os.unlink(filename)
def _dump(target, kind, suffix, content):
if target is not None:
print("====== {} DUMP ======".format(kind.upper()), file=sys.stderr)
content_value = content()
if isinstance(content_value, str):
content_value = bytes(content_value, 'utf-8')
if target == "":
file = tempfile.NamedTemporaryFile(suffix=suffix, delete=False)
else:
file = open(target + suffix, "wb")
file.write(content_value)
file.close()
print("{} dumped as {}".format(kind, file.name), file=sys.stderr)
class Target:
"""
A description of the target environment where the binaries
generated by the ARTIQ compiler will be deployed.
:var triple: (string)
LLVM target triple, e.g. ``"or1k"``
:var data_layout: (string)
LLVM target data layout, e.g. ``"E-m:e-p:32:32-i64:32-f64:32-v64:32-v128:32-a:0:32-n32"``
:var features: (list of string)
LLVM target CPU features, e.g. ``["mul", "div", "ffl1"]``
:var print_function: (string)
Name of a formatted print functions (with the signature of ``printf``)
provided by the target, e.g. ``"printf"``.
:var little_endian: (boolean)
Whether the code will be executed on a little-endian machine. This cannot be always
determined from data_layout due to JIT.
:var now_pinning: (boolean)
Whether the target implements the now-pinning RTIO optimization.
"""
triple = "unknown"
data_layout = ""
features = []
print_function = "printf"
little_endian = False
now_pinning = True
tool_ld = "ld.lld"
tool_strip = "llvm-strip"
tool_addr2line = "llvm-addr2line"
tool_cxxfilt = "llvm-cxxfilt"
def __init__(self):
self.llcontext = ll.Context()
def target_machine(self):
lltarget = llvm.Target.from_triple(self.triple)
llmachine = lltarget.create_target_machine(
features=",".join(["+{}".format(f) for f in self.features]),
reloc="pic", codemodel="default")
llmachine.set_asm_verbosity(True)
return llmachine
def optimize(self, llmodule):
llpassmgr = llvm.create_module_pass_manager()
# Register our alias analysis passes.
llpassmgr.add_basic_alias_analysis_pass()
llpassmgr.add_type_based_alias_analysis_pass()
# Start by cleaning up after our codegen and exposing as much
# information to LLVM as possible.
llpassmgr.add_constant_merge_pass()
llpassmgr.add_cfg_simplification_pass()
llpassmgr.add_instruction_combining_pass()
llpassmgr.add_sroa_pass()
llpassmgr.add_dead_code_elimination_pass()
llpassmgr.add_function_attrs_pass()
llpassmgr.add_global_optimizer_pass()
# Now, actually optimize the code.
llpassmgr.add_function_inlining_pass(275)
llpassmgr.add_ipsccp_pass()
llpassmgr.add_instruction_combining_pass()
llpassmgr.add_gvn_pass()
llpassmgr.add_cfg_simplification_pass()
llpassmgr.add_licm_pass()
# Clean up after optimizing.
llpassmgr.add_dead_arg_elimination_pass()
llpassmgr.add_global_dce_pass()
llpassmgr.run(llmodule)
def compile(self, module):
"""Compile the module to a relocatable object for this target."""
if os.getenv("ARTIQ_DUMP_SIG"):
print("====== MODULE_SIGNATURE DUMP ======", file=sys.stderr)
print(module, file=sys.stderr)
if os.getenv("ARTIQ_IR_NO_LOC") is not None:
ir.BasicBlock._dump_loc = False
type_printer = types.TypePrinter()
_dump(os.getenv("ARTIQ_DUMP_IR"), "ARTIQ IR", ".txt",
lambda: "\n".join(fn.as_entity(type_printer) for fn in module.artiq_ir))
llmod = module.build_llvm_ir(self)
try:
llparsedmod = llvm.parse_assembly(str(llmod))
llparsedmod.verify()
except RuntimeError:
_dump("", "LLVM IR (broken)", ".ll", lambda: str(llmod))
raise
_dump(os.getenv("ARTIQ_DUMP_UNOPT_LLVM"), "LLVM IR (generated)", "_unopt.ll",
lambda: str(llparsedmod))
self.optimize(llparsedmod)
_dump(os.getenv("ARTIQ_DUMP_LLVM"), "LLVM IR (optimized)", ".ll",
lambda: str(llparsedmod))
return llparsedmod
def assemble(self, llmodule):
llmachine = self.target_machine()
_dump(os.getenv("ARTIQ_DUMP_ASM"), "Assembly", ".s",
lambda: llmachine.emit_assembly(llmodule))
_dump(os.getenv("ARTIQ_DUMP_OBJ"), "Object file", ".o",
lambda: llmachine.emit_object(llmodule))
return llmachine.emit_object(llmodule)
def link(self, objects):
"""Link the relocatable objects into a shared library for this target."""
with RunTool([self.tool_ld, "-shared", "--eh-frame-hdr"] +
["{{obj{}}}".format(index) for index in range(len(objects))] +
["-x"] +
["-o", "{output}"],
output=None,
**{"obj{}".format(index): obj for index, obj in enumerate(objects)}) \
as results:
library = results["output"].read()
_dump(os.getenv("ARTIQ_DUMP_ELF"), "Shared library", ".elf",
lambda: library)
return library
def compile_and_link(self, modules):
return self.link([self.assemble(self.compile(module)) for module in modules])
def strip(self, library):
with RunTool([self.tool_strip, "--strip-debug", "{library}", "-o", "{output}"],
library=library, output=None) \
as results:
return results["output"].read()
def symbolize(self, library, addresses):
if addresses == []:
return []
# We got a list of return addresses, i.e. addresses of instructions
# just after the call. Offset them back to get an address somewhere
# inside the call instruction (or its delay slot), since that's what
# the backtrace entry should point at.
offset_addresses = [hex(addr - 1) for addr in addresses]
with RunTool([self.tool_addr2line, "--addresses", "--functions", "--inlines",
"--demangle", "--exe={library}"] + offset_addresses,
library=library) \
as results:
lines = iter(results["__stdout__"].read().rstrip().split("\n"))
backtrace = []
while True:
try:
address_or_function = next(lines)
except StopIteration:
break
if address_or_function[:2] == "0x":
address = int(address_or_function[2:], 16) + 1 # remove offset
function = next(lines)
else:
address = backtrace[-1][4] # inlined
function = address_or_function
location = next(lines)
filename, line = location.rsplit(":", 1)
if filename == "??" or filename == "<synthesized>":
continue
if line == "?":
line = -1
else:
line = int(line)
# can't get column out of addr2line D:
backtrace.append((filename, line, -1, function, address))
return backtrace
def demangle(self, names):
with RunTool([self.tool_cxxfilt] + names) as results:
return results["__stdout__"].read().rstrip().split("\n")
class NativeTarget(Target):
def __init__(self):
super().__init__()
self.triple = llvm.get_default_triple()
host_data_layout = str(llvm.targets.Target.from_default_triple().create_target_machine().target_data)
assert host_data_layout[0] in "eE"
self.little_endian = host_data_layout[0] == "e"
class OR1KTarget(Target):
triple = "or1k-linux"
data_layout = "E-m:e-p:32:32-i8:8:8-i16:16:16-i64:32:32-" \
"f64:32:32-v64:32:32-v128:32:32-a0:0:32-n32"
features = ["mul", "div", "ffl1", "cmov", "addc"]
print_function = "core_log"
little_endian = False
now_pinning = True
tool_ld = "or1k-linux-ld"
tool_strip = "or1k-linux-strip"
tool_addr2line = "or1k-linux-addr2line"
tool_cxxfilt = "or1k-linux-c++filt"
class CortexA9Target(Target):
triple = "armv7-unknown-linux-gnueabihf"
data_layout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
features = ["dsp", "fp16", "neon", "vfp3"]
print_function = "core_log"
little_endian = True
now_pinning = False
tool_ld = "armv7-unknown-linux-gnueabihf-ld"
tool_strip = "armv7-unknown-linux-gnueabihf-strip"
tool_addr2line = "armv7-unknown-linux-gnueabihf-addr2line"
tool_cxxfilt = "armv7-unknown-linux-gnueabihf-c++filt"