Poor performance? Slower than vtable? about proxy HOT 4 CLOSED

@ingangi I tried to compile your code with GCC 11.4, and observed that the virtual function calls in use_vtable() are devertualized, but not devertualized in use_proxy() and use_vtable_proxy(). See the screenshots below:

Indeed, GCC is better at devertualizing native virtual functions than proxy in this case. However, in most scenarios, runtime abstraction is used accross translation units or even ABI boundaries, and there is little chance for devirtualization. I think it would be more reasonable for the benchmark to separate the abstraction and implementation into different translation units, and see how it goes. From my personal experience, the numbers shoule be similar.

from proxy.

Update:
When I change the loop count from 10000 to 10, the result changed, proxy is faster.

from proxy.

@ingangi Could you share more context about the compiler version, flags and your test environment? Any template library, including STL implementations requires compiler optimizations to generate high quality code. If compiler optimization is properly set, I guess all the indirect calls in your code will be optimized away, because the implementation is in the same compilation unit. I suggest reviewing the generated code via Compiler Explorer.

from proxy.

@mingxwa thank you for the quick response.
compiler: g++ (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
flags: -std=c++20 -O2 -isystem benchmark/include -Lbenchmark/build/src -lbenchmark -lpthread

from proxy.