Files
cpython/Tools/cases_generator/generators_common.py
Ken Jin 4fa80ce74c gh-139109: A new tracing JIT compiler frontend for CPython (GH-140310)
This PR changes the current JIT model from trace projection to trace recording. Benchmarking: better pyperformance (about 1.7% overall) geomean versus current https://raw.githubusercontent.com/facebookexperimental/free-threading-benchmarking/refs/heads/main/results/bm-20251108-3.15.0a1%2B-7e2bc1d-JIT/bm-20251108-vultr-x86_64-Fidget%252dSpinner-tracing_jit-3.15.0a1%2B-7e2bc1d-vs-base.svg, 100% faster Richards on the most improved benchmark versus the current JIT. Slowdown of about 10-15% on the worst benchmark versus the current JIT. **Note: the fastest version isn't the one merged, as it relies on fixing bugs in the specializing interpreter, which is left to another PR**. The speedup in the merged version is about 1.1%. https://raw.githubusercontent.com/facebookexperimental/free-threading-benchmarking/refs/heads/main/results/bm-20251112-3.15.0a1%2B-f8a764a-JIT/bm-20251112-vultr-x86_64-Fidget%252dSpinner-tracing_jit-3.15.0a1%2B-f8a764a-vs-base.svg

Stats: 50% more uops executed, 30% more traces entered the last time we ran them. It also suggests our trace lengths for a real trace recording JIT are too short, as a lot of trace too long aborts https://github.com/facebookexperimental/free-threading-benchmarking/blob/main/results/bm-20251023-3.15.0a1%2B-eb73378-CLANG%2CJIT/bm-20251023-vultr-x86_64-Fidget%252dSpinner-tracing_jit-3.15.0a1%2B-eb73378-pystats-vs-base.md .

This new JIT frontend is already able to record/execute significantly more instructions than the previous JIT frontend. In this PR, we are now able to record through custom dunders, simple object creation, generators, etc. None of these were done by the old JIT frontend. Some custom dunders uops were discovered to be broken as part of this work gh-140277

The optimizer stack space check is disabled, as it's no longer valid to deal with underflow.

Pros:
* Ignoring the generated tracer code as it's automatically created, this is only additional 1k lines of code. The maintenance burden is handled by the DSL and code generator.
* `optimizer.c` is now significantly simpler, as we don't have to do strange things to recover the bytecode from a trace.
* The new JIT frontend is able to handle a lot more control-flow than the old one.
* Tracing is very low overhead. We use the tail calling interpreter/computed goto interpreter to switch between tracing mode and non-tracing mode. I call this mechanism dual dispatch, as we have two dispatch tables dispatching to each other. Specialization is still enabled while tracing.
* Better handling of polymorphism. We leverage the specializing interpreter for this.

Cons:
* (For now) requires tail calling interpreter or computed gotos. This means no Windows JIT for now :(. Not to fret, tail calling is coming soon to Windows though https://github.com/python/cpython/pull/139962

Design:
* After each instruction, the `record_previous_inst` function/label is executed. This does as the name suggests.
* The tracing interpreter lowers bytecode to uops directly so that it can obtain "fresh" values at the point of lowering.
* The tracing version behaves nearly identical to the normal interpreter, in fact it even has specialization! This allows it to run without much of a slowdown when tracing. The actual cost of tracing is only a function call and writes to memory.
* The tracing interpreter uses the specializing interpreter's deopt to naturally form the side exit chains. This allows it to side exit chain effectively, without repeating much code. We force a re-specializing when tracing a deopt.
* The tracing interpreter can even handle goto errors/exceptions, but I chose to disable them for now as it's not tested.
* Because we do not share interpreter dispatch, there is should be no significant slowdown to the original specializing interpreter on tailcall and computed got with JIT disabled. With JIT enabled, there might be a slowdown in the form of the JIT trying to trace.
* Things that could have dynamic instruction pointer effects are guarded on. The guard deopts to a new instruction --- `_DYNAMIC_EXIT`.
2025-11-13 18:08:32 +00:00

745 lines
23 KiB
Python

from pathlib import Path
from analyzer import (
Instruction,
Properties,
StackItem,
analysis_error,
Label,
CodeSection,
Uop,
)
from cwriter import CWriter
from typing import Callable, TextIO, Iterator, Iterable
from lexer import Token
from stack import Storage, StackError
from parser import Stmt, SimpleStmt, BlockStmt, IfStmt, ForStmt, WhileStmt, MacroIfStmt
from stack import PRINT_STACKS
DEBUG = False
class TokenIterator:
look_ahead: Token | None
iterator: Iterator[Token]
def __init__(self, tkns: Iterable[Token]):
self.iterator = iter(tkns)
self.look_ahead = None
def __iter__(self) -> "TokenIterator":
return self
def __next__(self) -> Token:
if self.look_ahead is None:
return next(self.iterator)
else:
res = self.look_ahead
self.look_ahead = None
return res
def peek(self) -> Token | None:
if self.look_ahead is None:
for tkn in self.iterator:
self.look_ahead = tkn
break
return self.look_ahead
ROOT = Path(__file__).parent.parent.parent.resolve()
DEFAULT_INPUT = (ROOT / "Python/bytecodes.c").as_posix()
def root_relative_path(filename: str) -> str:
try:
return Path(filename).resolve().relative_to(ROOT).as_posix()
except ValueError:
# Not relative to root, just return original path.
return filename
def type_and_null(var: StackItem) -> tuple[str, str]:
if var.is_array():
return "_PyStackRef *", "NULL"
else:
return "_PyStackRef", "PyStackRef_NULL"
def write_header(
generator: str, sources: list[str], outfile: TextIO, comment: str = "//"
) -> None:
outfile.write(
f"""{comment} This file is generated by {root_relative_path(generator)}
{comment} from:
{comment} {", ".join(root_relative_path(src) for src in sources)}
{comment} Do not edit!
"""
)
def emit_to(out: CWriter, tkn_iter: TokenIterator, end: str) -> Token:
parens = 0
for tkn in tkn_iter:
if tkn.kind == end and parens == 0:
return tkn
if tkn.kind == "LPAREN":
parens += 1
if tkn.kind == "RPAREN":
parens -= 1
out.emit(tkn)
raise analysis_error(f"Expecting {end}. Reached end of file", tkn)
ReplacementFunctionType = Callable[
[Token, TokenIterator, CodeSection, Storage, Instruction | None], bool
]
def always_true(tkn: Token | None) -> bool:
if tkn is None:
return False
return tkn.text in {"true", "1"}
NON_ESCAPING_DEALLOCS = {
"_PyFloat_ExactDealloc",
"_PyLong_ExactDealloc",
"_PyUnicode_ExactDealloc",
}
class Emitter:
out: CWriter
labels: dict[str, Label]
_replacers: dict[str, ReplacementFunctionType]
cannot_escape: bool
jump_prefix: str
def __init__(self, out: CWriter, labels: dict[str, Label], cannot_escape: bool = False, jump_prefix: str = ""):
self._replacers = {
"EXIT_IF": self.exit_if,
"AT_END_EXIT_IF": self.exit_if_after,
"DEOPT_IF": self.deopt_if,
"HANDLE_PENDING_AND_DEOPT_IF": self.periodic_if,
"ERROR_IF": self.error_if,
"ERROR_NO_POP": self.error_no_pop,
"DECREF_INPUTS": self.decref_inputs,
"DEAD": self.kill,
"INPUTS_DEAD": self.kill_inputs,
"SYNC_SP": self.sync_sp,
"SAVE_STACK": self.save_stack,
"RELOAD_STACK": self.reload_stack,
"PyStackRef_CLOSE_SPECIALIZED": self.stackref_close_specialized,
"PyStackRef_AsPyObjectSteal": self.stackref_steal,
"DISPATCH": self.dispatch,
"INSTRUCTION_SIZE": self.instruction_size,
"stack_pointer": self.stack_pointer,
}
self.out = out
self.labels = labels
self.cannot_escape = cannot_escape
self.jump_prefix = jump_prefix
def dispatch(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
if storage.spilled:
raise analysis_error("stack_pointer needs reloading before dispatch", tkn)
storage.stack.flush(self.out)
self.emit(tkn)
return False
def deopt_if(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
self.out.start_line()
self.out.emit("if (")
lparen = next(tkn_iter)
assert lparen.kind == "LPAREN"
first_tkn = tkn_iter.peek()
emit_to(self.out, tkn_iter, "RPAREN")
self.emit(") {\n")
next(tkn_iter) # Semi colon
assert inst is not None
assert inst.family is not None
family_name = inst.family.name
self.emit(f"UPDATE_MISS_STATS({family_name});\n")
self.emit(f"assert(_PyOpcode_Deopt[opcode] == ({family_name}));\n")
self.emit(f"JUMP_TO_PREDICTED({self.jump_prefix}{family_name});\n")
self.emit("}\n")
return not always_true(first_tkn)
exit_if = deopt_if
def periodic_if(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
raise NotImplementedError("HANDLE_PENDING_AND_DEOPT_IF not support in tier 1")
def exit_if_after(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
storage.clear_inputs("in AT_END_EXIT_IF")
storage.flush(self.out)
storage.stack.clear(self.out)
return self.exit_if(tkn, tkn_iter, uop, storage, inst)
def goto_error(self, offset: int, storage: Storage) -> str:
if offset > 0:
return f"{self.jump_prefix}JUMP_TO_LABEL(pop_{offset}_error);"
if offset < 0:
storage.copy().flush(self.out)
return f"{self.jump_prefix}JUMP_TO_LABEL(error);"
def error_if(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
lparen = next(tkn_iter)
assert lparen.kind == "LPAREN"
first_tkn = tkn_iter.peek()
unconditional = always_true(first_tkn)
if unconditional:
next(tkn_iter)
next(tkn_iter) # RPAREN
self.out.start_line()
else:
self.out.emit_at("if ", tkn)
self.emit(lparen)
emit_to(self.out, tkn_iter, "RPAREN")
self.out.emit(") {\n")
next(tkn_iter) # Semi colon
storage.clear_inputs("at ERROR_IF")
c_offset = storage.stack.sp_offset()
try:
offset = int(c_offset)
except ValueError:
offset = -1
self.out.emit(self.goto_error(offset, storage))
self.out.emit("\n")
if not unconditional:
self.out.emit("}\n")
return not unconditional
def error_no_pop(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter) # LPAREN
next(tkn_iter) # RPAREN
next(tkn_iter) # Semi colon
self.out.emit_at(self.goto_error(0, storage), tkn)
return False
def decref_inputs(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
next(tkn_iter)
next(tkn_iter)
self._print_storage("DECREF_INPUTS", storage)
try:
if not self.cannot_escape:
storage.close_inputs(self.out)
except StackError as ex:
raise analysis_error(ex.args[0], tkn)
except Exception as ex:
ex.args = (ex.args[0] + str(tkn),)
raise
return True
def kill_inputs(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
next(tkn_iter)
next(tkn_iter)
for var in storage.inputs:
var.kill()
return True
def kill(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
name_tkn = next(tkn_iter)
name = name_tkn.text
next(tkn_iter)
next(tkn_iter)
for var in storage.inputs:
if var.name == name:
var.kill()
break
else:
raise analysis_error(
f"'{name}' is not a live input-only variable", name_tkn
)
return True
def stackref_kill(
self,
name: Token,
storage: Storage,
escapes: bool
) -> bool:
live = ""
for var in reversed(storage.inputs):
if var.name == name.text:
if live and escapes:
raise analysis_error(
f"Cannot close '{name.text}' when "
f"'{live}' is still live", name)
var.kill()
break
if var.in_local:
live = var.name
return True
def stackref_close_specialized(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
self.out.emit(tkn)
tkn = next(tkn_iter)
assert tkn.kind == "LPAREN"
self.out.emit(tkn)
name = next(tkn_iter)
self.out.emit(name)
comma = next(tkn_iter)
if comma.kind != "COMMA":
raise analysis_error("Expected comma", comma)
self.out.emit(comma)
dealloc = next(tkn_iter)
if dealloc.kind != "IDENTIFIER":
raise analysis_error("Expected identifier", dealloc)
self.out.emit(dealloc)
if name.kind == "IDENTIFIER":
escapes = dealloc.text not in NON_ESCAPING_DEALLOCS
return self.stackref_kill(name, storage, escapes)
rparen = emit_to(self.out, tkn_iter, "RPAREN")
self.emit(rparen)
return True
def stackref_steal(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
self.out.emit(tkn)
tkn = next(tkn_iter)
assert tkn.kind == "LPAREN"
self.out.emit(tkn)
name = next(tkn_iter)
self.out.emit(name)
if name.kind == "IDENTIFIER":
return self.stackref_kill(name, storage, False)
rparen = emit_to(self.out, tkn_iter, "RPAREN")
self.emit(rparen)
return True
def sync_sp(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
next(tkn_iter)
next(tkn_iter)
storage.clear_inputs("when syncing stack")
storage.flush(self.out)
storage.stack.clear(self.out)
return True
def stack_pointer(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
if storage.spilled:
raise analysis_error("stack_pointer is invalid when stack is spilled to memory", tkn)
self.emit(tkn)
return True
def goto_label(self, goto: Token, label: Token, storage: Storage) -> None:
if label.text not in self.labels:
print(self.labels.keys())
raise analysis_error(f"Label '{label.text}' does not exist", label)
label_node = self.labels[label.text]
if label_node.spilled:
if not storage.spilled:
self.emit_save(storage)
elif storage.spilled:
raise analysis_error("Cannot jump from spilled label without reloading the stack pointer", goto)
self.out.start_line()
self.out.emit(f"{self.jump_prefix}JUMP_TO_LABEL(")
self.out.emit(label)
self.out.emit(")")
def emit_save(self, storage: Storage) -> None:
storage.flush(self.out)
storage.save(self.out)
def save_stack(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
next(tkn_iter)
next(tkn_iter)
self.emit_save(storage)
return True
def emit_reload(self, storage: Storage) -> None:
storage.reload(self.out)
def reload_stack(
self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
next(tkn_iter)
next(tkn_iter)
next(tkn_iter)
self.emit_reload(storage)
return True
def instruction_size(self,
tkn: Token,
tkn_iter: TokenIterator,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> bool:
"""Replace the INSTRUCTION_SIZE macro with the size of the current instruction."""
if uop.instruction_size is None:
raise analysis_error("The INSTRUCTION_SIZE macro requires uop.instruction_size to be set", tkn)
self.out.emit(f" {uop.instruction_size}u ")
return True
def _print_storage(self, reason:str, storage: Storage) -> None:
if DEBUG:
self.out.start_line()
self.emit(f"/* {reason} */\n")
self.emit(storage.as_comment())
self.out.start_line()
def _emit_stmt(
self,
stmt: Stmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
method_name = "emit_" + stmt.__class__.__name__
method = getattr(self, method_name, None)
if method is None:
raise NotImplementedError
return method(stmt, uop, storage, inst) # type: ignore[no-any-return]
def emit_SimpleStmt(
self,
stmt: SimpleStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
local_stores = set(uop.local_stores)
reachable = True
tkn = stmt.contents[-1]
try:
if stmt in uop.properties.escaping_calls and not self.cannot_escape:
escape = uop.properties.escaping_calls[stmt]
if escape.kills is not None:
self.stackref_kill(escape.kills, storage, True)
self.emit_save(storage)
tkn_iter = TokenIterator(stmt.contents)
for tkn in tkn_iter:
if tkn.kind == "GOTO":
label_tkn = next(tkn_iter)
self.goto_label(tkn, label_tkn, storage)
reachable = False
elif tkn.kind == "RETURN":
self.emit(tkn)
semicolon = emit_to(self.out, tkn_iter, "SEMI")
self.emit(semicolon)
reachable = False
elif tkn.kind == "IDENTIFIER":
if tkn.text in self._replacers:
if not self._replacers[tkn.text](tkn, tkn_iter, uop, storage, inst):
reachable = False
else:
if tkn in local_stores:
for var in storage.inputs:
if var.name == tkn.text:
var.in_local = True
var.memory_offset = None
break
for var in storage.outputs:
if var.name == tkn.text:
var.in_local = True
var.memory_offset = None
break
if tkn.text.startswith("DISPATCH"):
reachable = False
self.out.emit(tkn)
else:
self.out.emit(tkn)
if stmt in uop.properties.escaping_calls and not self.cannot_escape:
self.emit_reload(storage)
return reachable, None, storage
except StackError as ex:
raise analysis_error(ex.args[0], tkn) #from None
def emit_MacroIfStmt(
self,
stmt: MacroIfStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
self.out.emit(stmt.condition)
branch = stmt.else_ is not None
reachable = True
if branch:
else_storage = storage.copy()
for s in stmt.body:
r, tkn, storage = self._emit_stmt(s, uop, storage, inst)
if tkn is not None:
self.out.emit(tkn)
if not r:
reachable = False
if branch:
assert stmt.else_ is not None
self.out.emit(stmt.else_)
assert stmt.else_body is not None
for s in stmt.else_body:
r, tkn, else_storage = self._emit_stmt(s, uop, else_storage, inst)
if tkn is not None:
self.out.emit(tkn)
if not r:
reachable = False
else_storage.merge(storage, self.out) # type: ignore[possibly-undefined]
storage = else_storage
self.out.emit(stmt.endif)
return reachable, None, storage
def emit_IfStmt(
self,
stmt: IfStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
self.out.emit(stmt.if_)
for tkn in stmt.condition:
self.out.emit(tkn)
if_storage = storage.copy()
rbrace: Token | None = stmt.if_
try:
reachable, rbrace, if_storage = self._emit_stmt(stmt.body, uop, if_storage, inst)
if stmt.else_ is not None:
assert rbrace is not None
self.out.emit(rbrace)
self.out.emit(stmt.else_)
if stmt.else_body is not None:
else_reachable, rbrace, else_storage = self._emit_stmt(stmt.else_body, uop, storage, inst)
if not reachable:
reachable, storage = else_reachable, else_storage
elif not else_reachable:
# Discard the else storage
storage = if_storage
else:
#Both reachable
else_storage.merge(if_storage, self.out)
storage = else_storage
else:
if reachable:
if_storage.merge(storage, self.out)
storage = if_storage
else:
# Discard the if storage
reachable = True
return reachable, rbrace, storage
except StackError as ex:
assert rbrace is not None
raise analysis_error(ex.args[0], rbrace) from None
def emit_BlockStmt(
self,
stmt: BlockStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
emit_braces: bool = True,
) -> tuple[bool, Token | None, Storage]:
""" Returns (reachable?, closing '}', stack)."""
tkn: Token | None = None
try:
if emit_braces:
self.out.emit(stmt.open)
reachable = True
for s in stmt.body:
reachable, tkn, storage = self._emit_stmt(s, uop, storage, inst)
if tkn is not None:
self.out.emit(tkn)
if not reachable:
break
return reachable, stmt.close, storage
except StackError as ex:
if tkn is None:
tkn = stmt.close
raise analysis_error(ex.args[0], tkn) from None
def emit_ForStmt(
self,
stmt: ForStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
""" Returns (reachable?, closing '}', stack)."""
self.out.emit(stmt.for_)
for tkn in stmt.header:
self.out.emit(tkn)
return self._emit_stmt(stmt.body, uop, storage, inst)
def emit_WhileStmt(
self,
stmt: WhileStmt,
uop: CodeSection,
storage: Storage,
inst: Instruction | None,
) -> tuple[bool, Token | None, Storage]:
""" Returns (reachable?, closing '}', stack)."""
self.out.emit(stmt.while_)
for tkn in stmt.condition:
self.out.emit(tkn)
return self._emit_stmt(stmt.body, uop, storage, inst)
def emit_tokens(
self,
code: CodeSection,
storage: Storage,
inst: Instruction | None,
emit_braces: bool = True
) -> tuple[bool, Storage]:
self.out.start_line()
reachable, tkn, storage = self.emit_BlockStmt(code.body, code, storage, inst, emit_braces)
assert tkn is not None
try:
if reachable:
storage.push_outputs()
if emit_braces:
self.out.emit(tkn)
except StackError as ex:
raise analysis_error(ex.args[0], tkn) from None
return reachable, storage
def emit(self, txt: str | Token) -> None:
self.out.emit(txt)
def cflags(p: Properties) -> str:
flags: list[str] = []
if p.oparg:
flags.append("HAS_ARG_FLAG")
if p.uses_co_consts:
flags.append("HAS_CONST_FLAG")
if p.uses_co_names:
flags.append("HAS_NAME_FLAG")
if p.jumps:
flags.append("HAS_JUMP_FLAG")
if p.has_free:
flags.append("HAS_FREE_FLAG")
if p.uses_locals:
flags.append("HAS_LOCAL_FLAG")
if p.eval_breaker:
flags.append("HAS_EVAL_BREAK_FLAG")
if p.deopts:
flags.append("HAS_DEOPT_FLAG")
if p.deopts_periodic:
flags.append("HAS_PERIODIC_FLAG")
if p.side_exit:
flags.append("HAS_EXIT_FLAG")
if not p.infallible:
flags.append("HAS_ERROR_FLAG")
if p.error_without_pop:
flags.append("HAS_ERROR_NO_POP_FLAG")
if p.escapes:
flags.append("HAS_ESCAPES_FLAG")
if p.pure:
flags.append("HAS_PURE_FLAG")
if p.no_save_ip:
flags.append("HAS_NO_SAVE_IP_FLAG")
if p.unpredictable_jump:
flags.append("HAS_UNPREDICTABLE_JUMP_FLAG")
if p.needs_guard_ip:
flags.append("HAS_NEEDS_GUARD_IP_FLAG")
if flags:
return " | ".join(flags)
else:
return "0"