Run your benchmark for absl::AnyInvokable and faced with some issues #1851

MBkkt · 2025-03-05T04:31:59Z

MBkkt
Mar 5, 2025

clang 19, O3, libc++-19 abi v2, everything else is almost trunk (benchmark was extended to other function implementations)

Benchmark                                   Time             CPU   Iterations
-----------------------------------------------------------------------------
BM_TrivialStdFunction                    1.11 ns         1.11 ns    622220945
BM_TrivialAbslFunctionRef                1.10 ns         1.10 ns    637419510
BM_TrivialAbslAnyInvocable               1.85 ns         1.85 ns    377698611
BM_TrivialFollyFunctionRef               1.10 ns         1.10 ns    635925877
BM_TrivialFollyFunction                  2.03 ns         2.03 ns    345277532
BM_TrivialBoostFunction                  1.31 ns         1.31 ns    535099651
BM_TrivialFu2Function                    5.27 ns         5.27 ns    133167133
BM_TrivialFu2UniqueFunction              5.07 ns         5.07 ns    137751417

BM_LargeStdFunction                      11.0 ns         11.0 ns     63490131
BM_LargeAbslFunctionRef                  1.10 ns         1.10 ns    635590925
BM_LargeAbslAnyInvocable                 9.20 ns         9.20 ns     76082731
BM_LargeFollyFunctionRef                 1.10 ns         1.10 ns    635390697
BM_LargeFollyFunction                    11.1 ns         11.1 ns     62981528
BM_LargeBoostFunction                    11.9 ns         11.9 ns     58833683
BM_LargeFu2Function                      10.5 ns         10.5 ns     66229864
BM_LargeFu2UniqueFunction                10.7 ns         10.7 ns     65711158

BM_FunPtrStdFunction                     1.42 ns         1.42 ns    472854986
BM_FunPtrAbslFunctionRef                 1.28 ns         1.28 ns    545957263
BM_FunPtrAbslAnyInvocable                3.77 ns         3.77 ns    185835648
BM_FunPtrFollyFunctionRef                1.28 ns         1.28 ns    545872297
BM_FunPtrFollyFunction                   2.03 ns         2.03 ns    345335520
BM_FunPtrBoostFunction                   1.49 ns         1.49 ns    470280007
BM_FunPtrFu2Function                     5.98 ns         5.98 ns    117100264
BM_FunPtrFu2UniqueFunction               6.35 ns         6.35 ns    110392385

BM_TrivialArgsStdFunction               0.930 ns        0.930 ns    752811771
BM_TrivialArgsAbslFunctionRef           0.934 ns        0.934 ns    752592908
BM_TrivialArgsAbslAnyInvocable           1.11 ns         1.11 ns    630181056
BM_TrivialArgsFollyFunctionRef          0.952 ns        0.952 ns    743261325
BM_TrivialArgsFollyFunction             0.938 ns        0.938 ns    753110708
BM_TrivialArgsBoostFunction              1.12 ns         1.12 ns    627254290
BM_TrivialArgsFu2Function                2.24 ns         2.24 ns    312131551
BM_TrivialArgsFu2UniqueFunction          2.24 ns         2.24 ns    312115556

BM_NonTrivialArgsStdFunction             4.30 ns         4.30 ns    162679265
BM_NonTrivialArgsAbslFunctionRef         4.65 ns         4.65 ns    150336051
BM_NonTrivialArgsAbslAnyInvocable        4.29 ns         4.29 ns    162800037
BM_NonTrivialArgsFollyFunctionRef        4.65 ns         4.65 ns    150487034
BM_NonTrivialArgsFollyFunction           4.29 ns         4.29 ns    163183216
BM_NonTrivialArgsBoostFunction           7.48 ns         7.48 ns     93603218
BM_NonTrivialArgsFu2Function             5.94 ns         5.94 ns    117998045
BM_NonTrivialArgsFu2UniqueFunction       6.12 ns         6.12 ns    118745273

I interested why absl::AnyInvokable noticeably slower in case of function pointer
Maybe because of more expensive construction?
Because call-path looks even better except std::launder.
Also I didn't understand why does std::launder required for abseil implementation and doesn't for std and folly?
It doesn't affect microbenchmark