gh-146558: optimize dict access with objects with known hash

[kumaraditya303]

I asked Claude to create a benchmark for this, it created the following script which shows a real significant speedup.

Benchmark	No JIT	JIT	Speedup
dict_get[str]	55.7M/s	70.1M/s	+26%
dict_get[int]	55.5M/s	75.4M/s	+36%
dict_get[float]	50.9M/s	74.7M/s	+47%
dict_get[complex]	44.9M/s	74.9M/s	+67%
dict_store[str]	74.9M/s	85.7M/s	+14%
dict_store[int]	76.0M/s	94.1M/s	+24%
dict_store[float]	66.6M/s	97.0M/s	+46%
dict_store[complex]	55.9M/s	98.0M/s	+75%
dict_get[multi]	23.4M/s	35.1M/s	+50%
dict_get[obj]	52.9M/s	68.8M/s	+30%
dict_store[obj]	75.4M/s	92.5M/s	+23%

Details

"""Benchmark dict get/store with constant keys to measure_BINARY_OP_SUBSCR_DICT_KNOWN_HASH and _STORE_SUBSCR_DICT_KNOWN_HASH speedup.Usage: PYTHON_JIT=0 ./python.exe bench_dict_known_hash.py PYTHON_JIT=1 ./python.exe bench_dict_known_hash.py"""importtimeimportosimportsysN=50_000_000defbench_get_str(n):
d= {'a': 1, 'b': 2, 'c': 3}
x=0for_inrange(n):
x+=d['a']
returnxdefbench_get_int(n):
d= {1: 10, 2: 20, 3: 30}
x=0for_inrange(n):
x+=d[1]
returnxdefbench_store_str(n):
d= {}
for_inrange(n):
d['a'] =1returnddefbench_store_int(n):
d= {}
for_inrange(n):
d[1] =1returnddefbench_get_float(n):
d= {1.5: 1, 2.5: 2, 3.5: 3}
x=0for_inrange(n):
x+=d[1.5]
returnxdefbench_store_float(n):
d= {}
for_inrange(n):
d[1.5] =1returnddefbench_get_complex(n):
d= {1+2j: 1, 3+4j: 2}
x=0for_inrange(n):
x+=d[1+2j]
returnxdefbench_store_complex(n):
d= {}
for_inrange(n):
d[1+2j] =1returnddefbench_get_multi(n):
d= {'a': 1, 1: 2, b'x': 3, (1, 2): 4}
x=0for_inrange(n):
x+=d['a'] +d[1] +d[b'x'] +d[(1, 2)]
returnxclass_Key:
pass_KEY=_Key()
defbench_get_obj(n):
d= {_KEY: 1}
x=0for_inrange(n):
x+=d[_KEY]
returnxdefbench_store_obj(n):
d= {}
for_inrange(n):
d[_KEY] =1returnddefrun_avg(name, func, n, runs=3):
func(1000) # warmuptimes= []
for_inrange(runs):
t0=time.perf_counter()
func(n)
times.append(time.perf_counter() -t0)
avg=sum(times) /len(times)
rate=n/avg/1e6print(f" {name:25s}{avg:.3f}s ({rate:.1f}M iter/s)")
returnavgif__name__=="__main__":
jit="JIT"ifos.environ.get("PYTHON_JIT", "1") =="1"else"no JIT"print(f"Python {sys.version.split()[0]} ({jit}), avg of 3 runs\n")
run_avg("dict_get[str]", bench_get_str, N)
run_avg("dict_get[int]", bench_get_int, N)
run_avg("dict_get[float]", bench_get_float, N)
run_avg("dict_get[complex]", bench_get_complex, N)
run_avg("dict_store[str]", bench_store_str, N)
run_avg("dict_store[int]", bench_store_int, N)
run_avg("dict_store[float]", bench_store_float, N)
run_avg("dict_store[complex]",bench_store_complex,N)
run_avg("dict_get[multi]", bench_get_multi, N//4)
run_avg("dict_get[obj]", bench_get_obj, N)
run_avg("dict_store[obj]", bench_store_obj, N)

• Issue: JIT optimize dict access with consts with known hash #146558