Performance regression in scalar_mul with Clang and x86_64 assembly enabled #1682

1. 
   hebasto commented at 2:18 pm on June 5, 2025: member

   On Ubuntu 25.04:

   0$ ./build_clang17_0_6_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0319    ,     0.0345    ,     0.0569 
   4field_mul                     ,     0.0173    ,     0.0173    ,     0.0175
   5$ ./build_clang17_0_6_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0266    ,     0.0268    ,     0.0275 
   9field_mul                     ,     0.0173    ,     0.0174    ,     0.0177
   

   0$ ./build_clang18_1_8_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0336    ,     0.0345    ,     0.0357 
   4field_mul                     ,     0.0170    ,     0.0171    ,     0.0172
   5$ ./build_clang18_1_8_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0266    ,     0.0267    ,     0.0271 
   9field_mul                     ,     0.0170    ,     0.0171    ,     0.0174
   

   0$ ./build_clang19_1_7_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0329    ,     0.0330    ,     0.0331 
   4field_mul                     ,     0.0166    ,     0.0167    ,     0.0171 
   5$ ./build_clang19_1_7_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0270    ,     0.0270    ,     0.0271 
   9field_mul                     ,     0.0167    ,     0.0167    ,     0.0169 
   

   0$ ./build_clang20_1_2_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0330    ,     0.0343    ,     0.0447 
   4field_mul                     ,     0.0164    ,     0.0170    ,     0.0214 
   5$ ./build_clang20_1_2_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0269    ,     0.0270    ,     0.0271 
   9field_mul                     ,     0.0165    ,     0.0165    ,     0.0166 
   

2. 
   sipa commented at 2:19 pm on June 5, 2025: contributor

   Can you post benchmarks with GCC on the same machine?
3. 
   hebasto commented at 2:27 pm on June 5, 2025: member

   Can you post benchmarks with GCC on the same machine?

   Sure!

   0$ ./build_gcc11_5_0_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0289    ,     0.0291    ,     0.0301 
   4field_mul                     ,     0.0151    ,     0.0152    ,     0.0154 
   5$ ./build_gcc11_5_0_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0313    ,     0.0314    ,     0.0316 
   9field_mul                     ,     0.0151    ,     0.0151    ,     0.0153 
   

   0$ ./build_gcc15_0_1_asm/bin/bench_internal mul
   1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   2
   3scalar_mul                    ,     0.0292    ,     0.0293    ,     0.0295 
   4field_mul                     ,     0.0170    ,     0.0171    ,     0.0172
   5$ ./build_gcc15_0_1_noasm/bin/bench_internal mul
   6Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
   7
   8scalar_mul                    ,     0.0327    ,     0.0328    ,     0.0338 
   9field_mul                     ,     0.0171    ,     0.0171    ,     0.0172
   

4. real-or-random added the label performance on Jun 5, 2025
5. 
   real-or-random commented at 8:26 pm on June 5, 2025: contributor

   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.

Contributors
hebasto sipa real-or-random

Labels
performance