gh-146306: Optimize float operations by mutating uniquely-referenced operands in place (JIT only)

[eendebakpt]

We can add the following tier 2 micro-ops that mutate the uniquely-referenced operand:

• _BINARY_OP_ADD_FLOAT_INPLACE / _INPLACE_RIGHT — unique LHS / RHS
• _BINARY_OP_SUBTRACT_FLOAT_INPLACE / _INPLACE_RIGHT — unique LHS / RHS
• _BINARY_OP_MULTIPLY_FLOAT_INPLACE / _INPLACE_RIGHT — unique LHS / RHS
• _UNARY_NEGATIVE_FLOAT_INPLACE — unique operand

The _RIGHT variants handle commutative ops (add, multiply) plus subtract when only the RHS is unique. The optimizer emits these in optimizer_bytecodes.c when PyJitRef_IsUnique(left) or PyJitRef_IsUnique(right) is true and the operand is a known float. The mutated operand is marked as borrowed so the following _POP_TOP becomes _POP_TOP_NOP.

Micro-benchmarks:

Expression	main	optimized	Speedup
total += a*b + c	24.0 ns/iter	11.5 ns/iter	2.1x
total += a + b	16.9 ns/iter	11.0 ns/iter	1.5x
total += a*b + c*d	28.5 ns/iter	18.3 ns/iter	1.6x

pyperformance nbody (20k iterations):

	main	optimized	Speedup
nbody	60.6 ms	49.0 ms	1.19x (19% faster)

Followup

Some operations that could be added in followup PR's: division of floats (same as this PR, but needs to handle division by zero), operations on a float and an (compact) int with a uniquely referenced float, integer operations (but this is more involved because of small ints and number of digits), operations with complex numbers (same as this PR, but maybe the use case for complex is less).

Script

"""Demo script for the inplace float mutation optimization in the tier 2 optimizer.
./configure --enable-experimental-jit=interpreter --with-pydebug
./configure --enable-experimental-jit=yes --with-pydebug
Usage:
./python jit_float_demo.py # filtered trace (float-related ops)
./python jit_float_demo.py --all # full trace (all ops)
"""
import sys
import timeit
SHOW_ALL = "--all" in sys.argv
# --- System info ---
print("=" * 60)
print("CPython JIT / Tier 2 Demo")
print("=" * 60)
print(f"Python version: {sys.version}")
print(f"Debug build: {hasattr(sys, 'gettotalrefcount')}")
print(f"Free-threaded: {not sys._is_gil_enabled()}")
jit_mod = getattr(sys, "_jit", None)
if jit_mod is not None:
print(f"JIT available: {jit_mod.is_available()}")
print(f"JIT enabled: {jit_mod.is_enabled()}")
else:
print("JIT: not compiled in")
tier2 = False
try:
from _testinternalcapi import TIER2_THRESHOLD
tier2 = True
print(f"Tier 2: enabled (threshold={TIER2_THRESHOLD})")
except (ImportError, AttributeError):
print("Tier 2: disabled (build without _Py_TIER2)")
print()
# --- Example functions ---
def f_adds(n, a, b, c):
"""a*b + c per iteration — the multiply result is unique and reused."""
total = 0.0
for i in range(n):
total = a + b + c
return total
def f_chain(n, a, b, c):
"""a*b + c per iteration — the multiply result is unique and reused."""
total = 0.0
for i in range(n):
total += a * b + c
return total
def f_simple_add(n, a, b):
"""a + b per iteration — no unique intermediate."""
total = 0.0
for i in range(n):
total += a + b
return total
def f_long_chain(n, a, b, c, d):
"""a*b + c*d per iteration — two unique intermediates."""
total = 0.0
for i in range(n):
total += a * b + c * d
return total
def f_negate(n, a, b):
"""a*b + c*d per iteration — two unique intermediates."""
total = 0.0
for i in range(n):
total = - (a + b)
return total
# --- Warm up to trigger tier 2 ---
LOOP = 10_000
f_adds(LOOP, 2.0, 3.0, 4.0)
f_chain(LOOP, 2.0, 3.0, 4.0)
f_simple_add(LOOP, 2.0, 3.0)
f_long_chain(LOOP, 2.0, 3.0, 4.0, 5.0)
# --- Op annotation ---
def annotate_op(name, oparg, func):
"""Return a human-readable annotation for a uop."""
varnames = func.__code__.co_varnames
consts = func.__code__.co_consts
# _LOAD_FAST_BORROW_3 → local index 3
for prefix in ("_LOAD_FAST_BORROW_", "_LOAD_FAST_", "_SWAP_FAST_"):
if name.startswith(prefix):
idx = int(name[len(prefix):])
local = varnames[idx] if idx < len(varnames) else f"local{idx}"
return local
# _LOAD_CONST_INLINE_BORROW etc — operand is a pointer, not useful
# but if oparg is a small index into consts, show it
if "LOAD_CONST" in name and oparg < len(consts):
return repr(consts[oparg])
# Binary ops
if "MULTIPLY" in name:
return "*"
if "SUBTRACT" in name:
return "-"
if "ADD" in name and "UNICODE" not in name:
return "+"
# Guards
if name == "_GUARD_TOS_FLOAT":
return "top is float?"
if name == "_GUARD_NOS_FLOAT":
return "2nd is float?"
if name == "_GUARD_TOS_INT":
return "top is int?"
if name == "_GUARD_NOS_INT":
return "2nd is int?"
if "NOT_EXHAUSTED" in name:
return "iter not done?"
# Pop / cleanup
if name == "_POP_TOP_NOP":
return "skip (borrowed/null)"
if name == "_POP_TOP_FLOAT":
return "decref float"
if name == "_POP_TOP_INT":
return "decref int"
# Control flow
if name == "_JUMP_TO_TOP":
return "loop"
if name == "_EXIT_TRACE":
return "exit"
if name == "_DEOPT":
return "deoptimize"
if name == "_ERROR_POP_N":
return "error handler"
if name == "_START_EXECUTOR":
return "trace entry"
if name == "_MAKE_WARM":
return "warmup counter"
return ""
# --- Show traces ---
FILTER_KEYWORDS = (
"FLOAT", "INPLACE", "BINARY_OP", "NOP",
"LOAD_FAST", "LOAD_CONST", "GUARD",
)
has_get_executor = False
if tier2:
try:
from _opcode import get_executor
has_get_executor = True
except ImportError:
pass
if has_get_executor:
mode = "all ops" if SHOW_ALL else "float-related ops only"
print("-" * 60)
print(f"Tier 2 traces ({mode})")
print("-" * 60)
for label, func in [
("f_adds: total = a + b + c", f_adds),
("f_chain: total += a * b + c", f_chain),
("f_simple_add: total += a + b", f_simple_add),
("f_long_chain: total += a * b + c * d", f_long_chain),
]:
code = func.__code__
found = False
for i in range(len(code.co_code) // 2):
try:
ex = get_executor(code, i * 2)
except (ValueError, TypeError, RuntimeError):
continue
if ex is None:
continue
print(f"\n {label}")
for j, op in enumerate(ex):
name, oparg = op[0], op[1]
if not SHOW_ALL:
if not any(k in name for k in FILTER_KEYWORDS):
continue
annotation = annotate_op(name, oparg, func)
marker = " <<<" if "INPLACE" in name else ""
if annotation:
print(f" {j:3d}: {name:45s} # {annotation}{marker}")
else:
print(f" {j:3d}: {name}{marker}")
found = True
break
if not found:
print(f"\n {label}: (no executor found)")
print()
else:
print("-" * 60)
print("Tier 2 traces: skipped (tier 2 not available)")
print("-" * 60)
print()
# --- Benchmark ---
print("-" * 60)
print("Benchmark")
print("-" * 60)
N = 2_000_000
INNER = 1000
benchmarks = [
("total = a + b + c", lambda: f_adds(INNER, 2.0, 3.0, 4.0)),
("total += a*b + c ", lambda: f_chain(INNER, 2.0, 3.0, 4.0)),
("total += a + b ", lambda: f_simple_add(INNER, 2.0, 3.0)),
("total += a*b + c*d", lambda: f_long_chain(INNER, 2.0, 3.0, 4.0, 5.0)),
("total = - (a + b) ", lambda: f_negate(INNER, 2.0, 3.0)),
]
for label, fn in benchmarks:
iters = N // INNER
t = timeit.timeit(fn, number=iters)
ns_per = t / N * 1e9
print(f" {label}: {t:.3f}s ({ns_per:.0f} ns/iter)")
print()
print("The 'a*b + c' case benefits from _BINARY_OP_ADD_FLOAT_INPLACE:")
print("the result of a*b is uniquely referenced, so the addition")
print("mutates it in place instead of allocating a new float.")
# --- N-body benchmark from pyperformance ---
print()
print("-" * 60)
print("N-body benchmark (from pyperformance)")
print("-" * 60)
PI = 3.14159265358979323
SOLAR_MASS = 4 * PI * PI
DAYS_PER_YEAR = 365.24
def _nbody_make_system():
bodies = [
# sun
([0.0, 0.0, 0.0], [0.0, 0.0, 0.0], SOLAR_MASS),
# jupiter
([4.84143144246472090e+00, -1.16032004402742839e+00, -1.03622044471123109e-01],
[1.66007664274403694e-03*DAYS_PER_YEAR, 7.69901118419740425e-03*DAYS_PER_YEAR, -6.90460016972063023e-05*DAYS_PER_YEAR],
9.54791938424326609e-04*SOLAR_MASS),
# saturn
([8.34336671824457987e+00, 4.12479856412430479e+00, -4.03523417114321381e-01],
[-2.76742510726862411e-03*DAYS_PER_YEAR, 4.99852801234917238e-03*DAYS_PER_YEAR, 2.30417297573763929e-05*DAYS_PER_YEAR],
2.85885980666130812e-04*SOLAR_MASS),
# uranus
([1.28943695621391310e+01, -1.51111514016986312e+01, -2.23307578892655734e-01],
[2.96460137564761618e-03*DAYS_PER_YEAR, 2.37847173959480950e-03*DAYS_PER_YEAR, -2.96589568540237556e-05*DAYS_PER_YEAR],
4.36624404335156298e-05*SOLAR_MASS),
# neptune
([1.53796971148509165e+01, -2.59193146099879641e+01, 1.79258772950371181e-01],
[2.68067772490389322e-03*DAYS_PER_YEAR, 1.62824170038242295e-03*DAYS_PER_YEAR, -9.51592254519715870e-05*DAYS_PER_YEAR],
5.15138902046611451e-05*SOLAR_MASS),
]
pairs = []
for x in range(len(bodies) - 1):
for y in bodies[x + 1:]:
pairs.append((bodies[x], y))
return bodies, pairs
def _nbody_advance(dt, n, bodies, pairs):
for i in range(n):
for (([x1, y1, z1], v1, m1), ([x2, y2, z2], v2, m2)) in pairs:
dx = x1 - x2
dy = y1 - y2
dz = z1 - z2
mag = dt * ((dx * dx + dy * dy + dz * dz) ** (-1.5))
b1m = m1 * mag
b2m = m2 * mag
v1[0] -= dx * b2m
v1[1] -= dy * b2m
v1[2] -= dz * b2m
v2[0] += dx * b1m
v2[1] += dy * b1m
v2[2] += dz * b1m
for (r, [vx, vy, vz], m) in bodies:
r[0] += dt * vx
r[1] += dt * vy
r[2] += dt * vz
def _nbody_report_energy(bodies, pairs, e=0.0):
for (((x1, y1, z1), v1, m1), ((x2, y2, z2), v2, m2)) in pairs:
dx = x1 - x2
dy = y1 - y2
dz = z1 - z2
e -= (m1 * m2) / ((dx * dx + dy * dy + dz * dz) ** 0.5)
for (r, [vx, vy, vz], m) in bodies:
e += m * (vx * vx + vy * vy + vz * vz) / 2.
return e
def _nbody_offset_momentum(ref, bodies, px=0.0, py=0.0, pz=0.0):
for (r, [vx, vy, vz], m) in bodies:
px -= vx * m
py -= vy * m
pz -= vz * m
(r, v, m) = ref
v[0] = px / m
v[1] = py / m
v[2] = pz / m
def bench_nbody(iterations=20000):
bodies, pairs = _nbody_make_system()
_nbody_offset_momentum(bodies[0], bodies)
_nbody_report_energy(bodies, pairs)
_nbody_advance(0.01, iterations, bodies, pairs)
_nbody_report_energy(bodies, pairs)
# Warmup
bench_nbody(1000)
NBODY_RUNS = 5
t = timeit.timeit(lambda: bench_nbody(20000), number=NBODY_RUNS)
print(f" nbody (20k iterations, {NBODY_RUNS} runs): {t/NBODY_RUNS*1000:.1f} ms/run")

• Issue: Optimize float operations by mutating uniquely-referenced operands in place (JIT only) #146306

[eendebakpt]

Selected pyperformance benchmarks:

### chaos ###
Mean +- std dev: 60.7 ms +- 1.5 ms -> 59.5 ms +- 1.5 ms: 1.02x faster
Not significant
### fannkuch ###
Mean +- std dev: 376 ms +- 4 ms -> 375 ms +- 3 ms: 1.00x faster
Not significant
### float ###
Mean +- std dev: 60.9 ms +- 3.0 ms -> 58.8 ms +- 2.7 ms: 1.04x faster
Significant (t=4.00)
### nbody ###
Mean +- std dev: 87.2 ms +- 1.1 ms -> 72.3 ms +- 0.4 ms: 1.21x faster
Significant (t=98.34)
### raytrace ###
Mean +- std dev: 304 ms +- 5 ms -> 299 ms +- 5 ms: 1.02x faster
Not significant
### scimark_fft ###
Mean +- std dev: 305 ms +- 1 ms -> 297 ms +- 3 ms: 1.03x faster
Significant (t=19.01)
### scimark_lu ###
Mean +- std dev: 86.5 ms +- 0.4 ms -> 84.8 ms +- 0.4 ms: 1.02x faster
Not significant
### scimark_monte_carlo ###
Mean +- std dev: 60.7 ms +- 0.7 ms -> 59.8 ms +- 0.3 ms: 1.02x faster
Not significant
### scimark_sor ###
Mean +- std dev: 101 ms +- 1 ms -> 99 ms +- 3 ms: 1.03x faster
Significant (t=7.38)
### scimark_sparse_mat_mult ###
Mean +- std dev: 5.53 ms +- 0.02 ms -> 5.14 ms +- 0.02 ms: 1.08x faster
Significant (t=130.74)
### spectral_norm ###
Mean +- std dev: 84.0 ms +- 0.6 ms -> 79.0 ms +- 0.4 ms: 1.06x faster
Significant (t=53.73)

[Fidget-Spinner]

Isn't it more correct to say l = sym_new_null(ctx);? Same for below?

[eendebakpt]

To make this work I had to change the definition _POP_TOP_FLOAT. I think the change is ok, but please double check.

[Fidget-Spinner]

It should be PyJitRef_Borrow(sym_new_null(ctx)) sorry for not being clear

[Fidget-Spinner]

That way theres no need to change the case in POP_TOP_FLOAT

[eendebakpt]

Thanks for claryfying. Now _POP_TOP_FLOAT is unchanged.

[markshannon]

This looks like a nice performance improvement for floating point code
I've a few comments

[markshannon]

This op and its variants all share a lot of common code.
Could you factor out the code into a macro to perform the inplace operation?
Like:

tier2op(_BINARY_OP_ADD_FLOAT_INPLACE, (left, right--res, l, r)) {
res=FLOAT_INPLACE_OP(left, +, right);
l=PyStackRef_NULL;
r=right;
INPUTS_DEAD();
}
tier2op(_BINARY_OP_MULTIPLY_FLOAT_INPLACE_RIGHT, (left, right--res, l, r)) {
res=FLOAT_INPLACE_OP(right, *, left);
l=leftr=PyStackRef_NULL;
INPUTS_DEAD();
}

[eendebakpt]

Normal C macros are not allowed in the bytecodes.c opcodes (they are not expanded). I have not yet found a way to refactor this nicely.

We could add a new macro to the DSL (like INPUTS_DEAD) for this, but it feels a bit odd to add something to the DSL for this particular case.

[Fidget-Spinner]

Normal C macros are not allowed in the bytecodes.c opcodes

They are. You just need to define them in ceval_macros.h. Would you like me to do this, or would you like to do it?

[Fidget-Spinner]

Awesome. Thanks for doing this!

In the future, this will be superceded partially by the partial evaluation pass, as any Py_REFCNT(f) == 1 op is also safe to unbox. However, for 3.15 this is a good win.

[eendebakpt]

In the future, this will be superceded partially by the partial evaluation pass, as any Py_REFCNT(f) == 1 op is also safe to unbox. However, for 3.15 this is a good win.

Good to know! Given that, which of the operations listed in the followup section of the first post do you think would be candidates to add in new PRs?

[Fidget-Spinner]

Good to know! Given that, which of the operations listed in the followup section of the first post do you think would be candidates to add in new PRs?

I think compact_int OP compact_int is likely the next most common/worthwhile. The remaining mixed float-int stuff is less common in benchmarks.

They might be worthwhile, hard to tell.

[eendebakpt]

I checked compact int - compact int, but it is a bit messy: the operations can escape to non-compact ints or the result can be a small int. In these cases reusing the uniquely references is not straightforward.

The mixed float-int cases and float division are not too hard, and handling them has the advantage of propagating the type information. I will open new issues for these first.