Benchmark and assembly analysis gcc/llvm architecture specific optimization

Here I'm testing the differences between gcc/llvm with various optimization flags. Both performance and assembly are analyzed.

• To build ./build.sh
• To run ./run.sh
• To generate the asm ./generate.sh

The following combinations are tested:

• GCC
• GCC -O2
• GCC -O2 -march=native
• CLANG
• CLANG -O2
• CLANG -O2 -march=native

Test configurations

CPU: Intel(R) Core(TM) i7-7567U CPU @ 3.50GHz.
OS: Ubuntu 16-04 kernel 4.18.3 custom compiled.
main: dummy pi calculator.
gcc (Ubuntu 8.1.0-5ubuntu1~16.04) 8.1.0
clang version 7.0.0 (trunk 338359) (llvm/trunk 338357)

Results

iterations 1000000000 value 3.14159

• clang time 5.70893s
• gcc time 9.08183s
• clang -O2 time 6.01251s
• gcc -O2 time 5.13561s
• clang -march=native -O2 time 1.21312s
• gcc -march=native -O2 time 4.7192s

References

• http://lists.llvm.org/pipermail/cfe-dev/2016-March/048049.html
• http://sdf.org/~riley/blog/2014/10/30/march-mtune/
• https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html