bpo-47053: Reduce deoptimization in BINARY_OP_INPLACE_ADD_UNICODE #31318
Conversation
|
A benchmark: from pyperf import Runner, perf_counter
import sys
LENGTHS = [3, 1, 4, 1, 5, 9, 2, 6, 5, 3, 5, 8, 9]
DATA = [('a',) * x for x in LENGTHS]
def bench(loops):
data = DATA * loops
t0 = perf_counter()
for s in data:
res = ''
for c in s:
res += c
return perf_counter() - t0
runner = Runner()
runner.bench_time_func("inplace add str", bench) |
|
Looks good for the microbenchmark, do you have numbers for the full suite? This is the stats for this PR? |
On my not-that-stable laptop with GCC on WSL, using --enable-optimizations --with-lto, I get a 1.02x faster geometric mean: https://gist.github.com/sweeneyde/7fb779d28c55ba4b5e8d40f0bf8f596f
Yes, that's correct. |
|
The previous microbenchmark was with MSCV, I got some different results with related benchmarks on GCC: from pyperf import Runner, perf_counter
import sys
LENGTHS = [3, 1, 4, 1, 5, 9, 2, 6, 5, 3, 5, 8, 9]
DATA = [('a',) * x for x in LENGTHS]
def bench1(loops):
data = DATA * loops
t0 = perf_counter()
for s in data:
res = ''
for c in s:
res += c
return perf_counter() - t0
def bench2(loops):
data = []
n = 10_000
for i in range(loops):
data.append(("a" * n, "a"))
data.append(("", "a"))
t0 = perf_counter()
for s, c in data:
s += c
s += c
s += c
return perf_counter() - t0
runner = Runner()
runner.bench_time_func("bench1", bench1)
runner.bench_time_func("bench2", bench2)It could be just a random result of how the PGO decides to shuffle things around? |
|
~0% misses is a very clear improvement on 96.6%. I'm not worried that the benchmarks numbers are a bit noisy, given this is a clear improvement. |
Hopefully "left-hand side is the same as assignment target" is more stable and less miss-prone than
Py_REFCNT == 2.Note that PyUnicode_Append already has lots of overhead, and it checks if it's safe to work in place.
See faster-cpython/ideas#269
https://bugs.python.org/issue47053