Files
cpython/Tools/cases_generator/opcode_metadata_generator.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The optimizer stack space check is disabled, as it's no longer valid to deal with underflow.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

416 lines
14 KiB
Python

"""Generate opcode metadata.
Reads the instruction definitions from bytecodes.c.
Writes the metadata to pycore_opcode_metadata.h by default.
"""
import argparse
from analyzer import (
Analysis,
Instruction,
PseudoInstruction,
analyze_files,
Uop,
)
from generators_common import (
DEFAULT_INPUT,
ROOT,
write_header,
cflags,
)
from cwriter import CWriter
from dataclasses import dataclass
from typing import TextIO
from stack import get_stack_effect
# Constants used instead of size for macro expansions.
# Note: 1, 2, 4 must match actual cache entry sizes.
OPARG_KINDS = {
"OPARG_SIMPLE": 0,
"OPARG_CACHE_1": 1,
"OPARG_CACHE_2": 2,
"OPARG_CACHE_4": 4,
"OPARG_TOP": 5,
"OPARG_BOTTOM": 6,
"OPARG_SAVE_RETURN_OFFSET": 7,
# Skip 8 as the other powers of 2 are sizes
"OPARG_REPLACED": 9,
"OPERAND1_1": 10,
"OPERAND1_2": 11,
"OPERAND1_4": 12,
}
FLAGS = [
"ARG",
"CONST",
"NAME",
"JUMP",
"FREE",
"LOCAL",
"EVAL_BREAK",
"DEOPT",
"ERROR",
"ESCAPES",
"EXIT",
"PURE",
"ERROR_NO_POP",
"NO_SAVE_IP",
"PERIODIC",
"UNPREDICTABLE_JUMP",
"NEEDS_GUARD_IP",
]
def generate_flag_macros(out: CWriter) -> None:
for i, flag in enumerate(FLAGS):
out.emit(f"#define HAS_{flag}_FLAG ({1<<i})\n")
for i, flag in enumerate(FLAGS):
out.emit(
f"#define OPCODE_HAS_{flag}(OP) (_PyOpcode_opcode_metadata[OP].flags & (HAS_{flag}_FLAG))\n"
)
out.emit("\n")
def generate_oparg_macros(out: CWriter) -> None:
for name, value in OPARG_KINDS.items():
out.emit(f"#define {name} {value}\n")
out.emit("\n")
def emit_stack_effect_function(
out: CWriter, direction: str, data: list[tuple[str, str]]
) -> None:
out.emit(f"extern int _PyOpcode_num_{direction}(int opcode, int oparg);\n")
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit(f"int _PyOpcode_num_{direction}(int opcode, int oparg) {{\n")
out.emit("switch(opcode) {\n")
for name, effect in data:
out.emit(f"case {name}:\n")
out.emit(f" return {effect};\n")
out.emit("default:\n")
out.emit(" return -1;\n")
out.emit("}\n")
out.emit("}\n\n")
out.emit("#endif\n\n")
def generate_stack_effect_functions(analysis: Analysis, out: CWriter) -> None:
popped_data: list[tuple[str, str]] = []
pushed_data: list[tuple[str, str]] = []
def add(inst: Instruction | PseudoInstruction) -> None:
stack = get_stack_effect(inst)
popped = (-stack.base_offset).to_c()
pushed = (stack.logical_sp - stack.base_offset).to_c()
popped_data.append((inst.name, popped))
pushed_data.append((inst.name, pushed))
for inst in analysis.instructions.values():
add(inst)
for pseudo in analysis.pseudos.values():
add(pseudo)
emit_stack_effect_function(out, "popped", sorted(popped_data))
emit_stack_effect_function(out, "pushed", sorted(pushed_data))
def generate_is_pseudo(analysis: Analysis, out: CWriter) -> None:
"""Write the IS_PSEUDO_INSTR macro"""
out.emit("\n\n#define IS_PSEUDO_INSTR(OP) ( \\\n")
for op in analysis.pseudos:
out.emit(f"((OP) == {op}) || \\\n")
out.emit("0")
out.emit(")\n\n")
def get_format(inst: Instruction) -> str:
if inst.properties.oparg:
format = "INSTR_FMT_IB"
else:
format = "INSTR_FMT_IX"
if inst.size > 1:
format += "C"
format += "0" * (inst.size - 2)
return format
def generate_instruction_formats(analysis: Analysis, out: CWriter) -> None:
# Compute the set of all instruction formats.
formats: set[str] = set()
for inst in analysis.instructions.values():
formats.add(get_format(inst))
# Generate an enum for it
out.emit("enum InstructionFormat {\n")
next_id = 1
for format in sorted(formats):
out.emit(f"{format} = {next_id},\n")
next_id += 1
out.emit("};\n\n")
def generate_deopt_table(analysis: Analysis, out: CWriter) -> None:
out.emit("extern const uint8_t _PyOpcode_Deopt[256];\n")
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit("const uint8_t _PyOpcode_Deopt[256] = {\n")
deopts: list[tuple[str, str]] = []
for inst in analysis.instructions.values():
deopt = inst.name
if inst.family is not None:
deopt = inst.family.name
deopts.append((inst.name, deopt))
defined = set(analysis.opmap.values())
for i in range(256):
if i not in defined:
deopts.append((f'{i}', f'{i}'))
assert len(deopts) == 256
assert len(set(x[0] for x in deopts)) == 256
for name, deopt in sorted(deopts):
out.emit(f"[{name}] = {deopt},\n")
out.emit("};\n\n")
out.emit("#endif // NEED_OPCODE_METADATA\n\n")
def generate_cache_table(analysis: Analysis, out: CWriter) -> None:
out.emit("extern const uint8_t _PyOpcode_Caches[256];\n")
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit("const uint8_t _PyOpcode_Caches[256] = {\n")
for inst in analysis.instructions.values():
if inst.family and inst.family.name != inst.name:
continue
if inst.name.startswith("INSTRUMENTED"):
continue
if inst.size > 1:
out.emit(f"[{inst.name}] = {inst.size-1},\n")
out.emit("};\n")
out.emit("#endif\n\n")
def generate_name_table(analysis: Analysis, out: CWriter) -> None:
table_size = 256 + len(analysis.pseudos)
out.emit(f"extern const char *_PyOpcode_OpName[{table_size}];\n")
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit(f"const char *_PyOpcode_OpName[{table_size}] = {{\n")
names = list(analysis.instructions) + list(analysis.pseudos)
for name in sorted(names):
out.emit(f'[{name}] = "{name}",\n')
out.emit("};\n")
out.emit("#endif\n\n")
def generate_metadata_table(analysis: Analysis, out: CWriter) -> None:
table_size = 256 + len(analysis.pseudos)
out.emit("struct opcode_metadata {\n")
out.emit("uint8_t valid_entry;\n")
out.emit("uint8_t instr_format;\n")
out.emit("uint32_t flags;\n")
out.emit("};\n\n")
out.emit(
f"extern const struct opcode_metadata _PyOpcode_opcode_metadata[{table_size}];\n"
)
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit(
f"const struct opcode_metadata _PyOpcode_opcode_metadata[{table_size}] = {{\n"
)
for inst in sorted(analysis.instructions.values(), key=lambda t: t.name):
out.emit(
f"[{inst.name}] = {{ true, {get_format(inst)}, {cflags(inst.properties)} }},\n"
)
for pseudo in sorted(analysis.pseudos.values(), key=lambda t: t.name):
flags = cflags(pseudo.properties)
for flag in pseudo.flags:
if flags == "0":
flags = f"{flag}_FLAG"
else:
flags += f" | {flag}_FLAG"
out.emit(f"[{pseudo.name}] = {{ true, -1, {flags} }},\n")
out.emit("};\n")
out.emit("#endif\n\n")
def generate_expansion_table(analysis: Analysis, out: CWriter) -> None:
expansions_table: dict[str, list[tuple[str, str, int]]] = {}
for inst in sorted(analysis.instructions.values(), key=lambda t: t.name):
offset: int = 0 # Cache effect offset
expansions: list[tuple[str, str, int]] = [] # [(name, size, offset), ...]
if inst.is_super():
pieces = inst.name.split("_")
assert len(pieces) % 2 == 0, f"{inst.name} doesn't look like a super-instr"
parts_per_piece = int(len(pieces) / 2)
name1 = "_".join(pieces[:parts_per_piece])
name2 = "_".join(pieces[parts_per_piece:])
assert name1 in analysis.instructions, f"{name1} doesn't match any instr"
assert name2 in analysis.instructions, f"{name2} doesn't match any instr"
instr1 = analysis.instructions[name1]
instr2 = analysis.instructions[name2]
for part in instr1.parts:
expansions.append((part.name, "OPARG_TOP", 0))
for part in instr2.parts:
expansions.append((part.name, "OPARG_BOTTOM", 0))
elif not is_viable_expansion(inst):
continue
else:
for part in inst.parts:
size = part.size
if isinstance(part, Uop):
# Skip specializations
if "specializing" in part.annotations:
continue
# Add the primary expansion.
fmt = "OPARG_SIMPLE"
if part.name == "_SAVE_RETURN_OFFSET":
fmt = "OPARG_SAVE_RETURN_OFFSET"
elif part.caches:
fmt = str(part.caches[0].size)
if "replaced" in part.annotations:
fmt = "OPARG_REPLACED"
expansions.append((part.name, fmt, offset))
if len(part.caches) > 1:
# Add expansion for the second operand
internal_offset = 0
for cache in part.caches[:-1]:
internal_offset += cache.size
expansions.append((part.name, f"OPERAND1_{part.caches[-1].size}", offset+internal_offset))
offset += part.size
expansions_table[inst.name] = expansions
max_uops = max(len(ex) for ex in expansions_table.values())
out.emit(f"#define MAX_UOP_PER_EXPANSION {max_uops}\n")
out.emit("struct opcode_macro_expansion {\n")
out.emit("int nuops;\n")
out.emit(
"struct { int16_t uop; int8_t size; int8_t offset; } uops[MAX_UOP_PER_EXPANSION];\n"
)
out.emit("};\n")
out.emit(
"extern const struct opcode_macro_expansion _PyOpcode_macro_expansion[256];\n\n"
)
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit("const struct opcode_macro_expansion\n")
out.emit("_PyOpcode_macro_expansion[256] = {\n")
for inst_name, expansions in expansions_table.items():
uops = [
f"{{ {name}, {size}, {offset} }}" for (name, size, offset) in expansions
]
out.emit(
f'[{inst_name}] = {{ .nuops = {len(expansions)}, .uops = {{ {", ".join(uops)} }} }},\n'
)
out.emit("};\n")
out.emit("#endif // NEED_OPCODE_METADATA\n\n")
def is_viable_expansion(inst: Instruction) -> bool:
"An instruction can be expanded if all its parts are viable for tier 2"
for part in inst.parts:
if isinstance(part, Uop):
# Skip specializing and replaced uops
if "specializing" in part.annotations:
continue
if "replaced" in part.annotations:
continue
if part.properties.tier == 1 or not part.is_viable():
return False
return True
def generate_extra_cases(analysis: Analysis, out: CWriter) -> None:
out.emit("#define EXTRA_CASES \\\n")
valid_opcodes = set(analysis.opmap.values())
for op in range(256):
if op not in valid_opcodes:
out.emit(f" case {op}: \\\n")
out.emit(" ;\n")
def generate_pseudo_targets(analysis: Analysis, out: CWriter) -> None:
table_size = len(analysis.pseudos)
max_targets = max(len(pseudo.targets) for pseudo in analysis.pseudos.values())
out.emit("struct pseudo_targets {\n")
out.emit(f"uint8_t as_sequence;\n")
out.emit(f"uint8_t targets[{max_targets + 1}];\n")
out.emit("};\n")
out.emit(
f"extern const struct pseudo_targets _PyOpcode_PseudoTargets[{table_size}];\n"
)
out.emit("#ifdef NEED_OPCODE_METADATA\n")
out.emit(
f"const struct pseudo_targets _PyOpcode_PseudoTargets[{table_size}] = {{\n"
)
for pseudo in analysis.pseudos.values():
as_sequence = "1" if pseudo.as_sequence else "0"
targets = ["0"] * (max_targets + 1)
for i, target in enumerate(pseudo.targets):
targets[i] = target.name
out.emit(f"[{pseudo.name}-256] = {{ {as_sequence}, {{ {', '.join(targets)} }} }},\n")
out.emit("};\n\n")
out.emit("#endif // NEED_OPCODE_METADATA\n")
out.emit("static inline bool\n")
out.emit("is_pseudo_target(int pseudo, int target) {\n")
out.emit(f"if (pseudo < 256 || pseudo >= {256+table_size}) {{\n")
out.emit(f"return false;\n")
out.emit("}\n")
out.emit(
f"for (int i = 0; _PyOpcode_PseudoTargets[pseudo-256].targets[i]; i++) {{\n"
)
out.emit(
f"if (_PyOpcode_PseudoTargets[pseudo-256].targets[i] == target) return true;\n"
)
out.emit("}\n")
out.emit(f"return false;\n")
out.emit("}\n\n")
def generate_opcode_metadata(
filenames: list[str], analysis: Analysis, outfile: TextIO
) -> None:
write_header(__file__, filenames, outfile)
out = CWriter(outfile, 0, False)
with out.header_guard("Py_CORE_OPCODE_METADATA_H"):
out.emit("#ifndef Py_BUILD_CORE\n")
out.emit('# error "this header requires Py_BUILD_CORE define"\n')
out.emit("#endif\n\n")
out.emit("#include <stdbool.h> // bool\n")
out.emit('#include "opcode_ids.h"\n')
generate_is_pseudo(analysis, out)
out.emit('#include "pycore_uop_ids.h"\n')
generate_stack_effect_functions(analysis, out)
generate_instruction_formats(analysis, out)
table_size = 256 + len(analysis.pseudos)
out.emit("#define IS_VALID_OPCODE(OP) \\\n")
out.emit(f" (((OP) >= 0) && ((OP) < {table_size}) && \\\n")
out.emit(" (_PyOpcode_opcode_metadata[(OP)].valid_entry))\n\n")
generate_flag_macros(out)
generate_oparg_macros(out)
generate_metadata_table(analysis, out)
generate_expansion_table(analysis, out)
generate_name_table(analysis, out)
generate_cache_table(analysis, out)
generate_deopt_table(analysis, out)
generate_extra_cases(analysis, out)
generate_pseudo_targets(analysis, out)
arg_parser = argparse.ArgumentParser(
description="Generate the header file with opcode metadata.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
DEFAULT_OUTPUT = ROOT / "Include/internal/pycore_opcode_metadata.h"
arg_parser.add_argument(
"-o", "--output", type=str, help="Generated code", default=DEFAULT_OUTPUT
)
arg_parser.add_argument(
"input", nargs=argparse.REMAINDER, help="Instruction definition file(s)"
)
if __name__ == "__main__":
args = arg_parser.parse_args()
if len(args.input) == 0:
args.input.append(DEFAULT_INPUT)
data = analyze_files(args.input)
with open(args.output, "w") as outfile:
generate_opcode_metadata(args.input, data, outfile)