Clang beats handwritten `x86_64` assembly for `scalar_mul` #1682

On Ubuntu 25.04:

$ ./build_clang17_0_6_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0319    ,     0.0345    ,     0.0569 
field_mul                     ,     0.0173    ,     0.0173    ,     0.0175
$ ./build_clang17_0_6_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0266    ,     0.0268    ,     0.0275 
field_mul                     ,     0.0173    ,     0.0174    ,     0.0177

$ ./build_clang18_1_8_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0336    ,     0.0345    ,     0.0357 
field_mul                     ,     0.0170    ,     0.0171    ,     0.0172
$ ./build_clang18_1_8_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0266    ,     0.0267    ,     0.0271 
field_mul                     ,     0.0170    ,     0.0171    ,     0.0174

$ ./build_clang19_1_7_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0329    ,     0.0330    ,     0.0331 
field_mul                     ,     0.0166    ,     0.0167    ,     0.0171 
$ ./build_clang19_1_7_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0270    ,     0.0270    ,     0.0271 
field_mul                     ,     0.0167    ,     0.0167    ,     0.0169 

$ ./build_clang20_1_2_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0330    ,     0.0343    ,     0.0447 
field_mul                     ,     0.0164    ,     0.0170    ,     0.0214 
$ ./build_clang20_1_2_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0269    ,     0.0270    ,     0.0271 
field_mul                     ,     0.0165    ,     0.0165    ,     0.0166 

Can you post benchmarks with GCC on the same machine?

Can you post benchmarks with GCC on the same machine?

Sure!

$ ./build_gcc11_5_0_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0289    ,     0.0291    ,     0.0301 
field_mul                     ,     0.0151    ,     0.0152    ,     0.0154 
$ ./build_gcc11_5_0_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0313    ,     0.0314    ,     0.0316 
field_mul                     ,     0.0151    ,     0.0151    ,     0.0153 

$ ./build_gcc15_0_1_asm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0292    ,     0.0293    ,     0.0295 
field_mul                     ,     0.0170    ,     0.0171    ,     0.0172
$ ./build_gcc15_0_1_noasm/bin/bench_internal mul
Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    

scalar_mul                    ,     0.0327    ,     0.0328    ,     0.0338 
field_mul                     ,     0.0171    ,     0.0171    ,     0.0172

Fwiw, I've hacked together a compiler explorer instance where you can compare clang's output (inline asm disabled) on different versions. I don't see a big change, it's probably just a lot of incremental improvements that add up.

Am I right that there's no "regression" in the normal meaning of the word here (i.e., something which was fast on previous library or compiler versions became slower)? It's just that the Clang's code became faster, and it's faster than the ASM. (Let me edit the title.)