Investigate why build artifacts are significantly slower with gcc than with clang

[mhx]

No description provided.

[mhx]

Found the problem: gcc -O3 is a bad idea. Funnily enough, I was aware that -O3 can cause problems, and that it's for some reason CMake's default. However, I never seemed to connect the dots until now.

mhx@sauron ~ $ hyperfine -L prog git/github/dwarfs/build-clang-11/mkdwarfs,git/github/dwarfs/build-clang-11-O2/mkdwarfs,git/github/dwarfs/build-gcc-10/mkdwarfs,git/github/dwarfs/build-gcc-10-O2/mkdwarfs '{prog} -i install-medium -o /dev
/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12'                                                                                                                                                                           
Benchmark #1: git/github/dwarfs/build-clang-11/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12                                                                                           
  Time (mean ± σ):     14.460 s ±  0.335 s    [User: 85.348 s, System: 13.218 s]                                                                                                                                                            
  Range (min … max):   14.001 s … 14.996 s    10 runs                                                                                                                                                                                       
                                                                                                                                                                                                                                            
Benchmark #2: git/github/dwarfs/build-clang-11-O2/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12                                                                              
  Time (mean ± σ):     14.764 s ±  0.177 s    [User: 88.102 s, System: 13.452 s]
  Range (min … max):   14.444 s … 15.117 s    10 runs
  
Benchmark #3: git/github/dwarfs/build-gcc-10/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12
  Time (mean ± σ):     20.201 s ±  0.126 s    [User: 93.495 s, System: 13.381 s]
  Range (min … max):   20.087 s … 20.487 s    10 runs
  
Benchmark #4: git/github/dwarfs/build-gcc-10-O2/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12
  Time (mean ± σ):     15.253 s ±  0.162 s    [User: 89.495 s, System: 13.316 s]
  Range (min … max):   14.915 s … 15.522 s    10 runs
  
Summary
  'git/github/dwarfs/build-clang-11/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12' ran
    1.02 ± 0.03 times faster than 'git/github/dwarfs/build-clang-11-O2/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12'
    1.05 ± 0.03 times faster than 'git/github/dwarfs/build-gcc-10-O2/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12'
    1.40 ± 0.03 times faster than 'git/github/dwarfs/build-gcc-10/mkdwarfs -i install-medium -o /dev/null -l0 --order=nilsimsa:255:20000:20000 -S24 --window-size=12'

There's some extremely hot code in one particular function that gcc -O3 messes up completely. It's not that clang is without flaws [1], but in general it seems to do a reasonable job with -O3 (and it generates faster code than gcc anyway).

[1] When given the chance to choose between popcount and a set of AVX instruction, clang picks AVX, which turns out to be a lot slower than popcount.

[mhx]

Fixed in dwarfs-0.4.1.