[llvm] [RISCV] Enable select optimization by default (PR #178394)

[Craig Topper via llvm-commits]

topperc wrote:

> Results of compiling`CTMark` (-n is an option to filter benches by regexes):
> 
> ```
> ../utils/compare.py off.json on.json -n .*CTMark.* -m compile_time 
> Tests: 2095
> In filter names: 2085 (filtered out)
> Remaining: 10
> Metric: compile_time
> 
> Program                                       compile_time             
>                                               off          on     diff 
> sqlite3/sqlite3                                17.04        18.24  7.0%
> SPASS/SPASS                                    28.98        29.50  1.8%
> mafft/pairlocalalign                           17.09        17.31  1.3%
> tramp3d-v4/tramp3d-v4                          44.07        44.45  0.9%
> ClamAV/clamscan                                32.94        32.80 -0.4%
> 7zip/7zip-benchmark                           146.36       145.02 -0.9%
> lencod/lencod                                  38.00        37.38 -1.6%
> consumer-typeset/consumer-typeset              22.75        21.96 -3.5%
> kimwitu++/kc                                   29.61        28.18 -4.8%
> Bullet/bullet                                  73.31        68.29 -6.8%
>                            Geomean difference                     -0.8%
> ```
> 
> ```
> ../utils/compare.py off.json on.json -n .*CTMark.* -m compile_time 
> Tests: 2095
> In filter names: 2085 (filtered out)
> Remaining: 10
> Metric: compile_time
> 
> Program                                       compile_time              
>                                               off          on     diff  
> 7zip/7zip-benchmark                           143.58       142.26  -0.9%
> mafft/pairlocalalign                           16.55        16.33  -1.3%
> sqlite3/sqlite3                                17.26        16.48  -4.5%
> consumer-typeset/consumer-typeset              20.98        19.75  -5.8%
> Bullet/bullet                                  69.76        65.62  -5.9%
> tramp3d-v4/tramp3d-v4                          46.07        42.69  -7.3%
> lencod/lencod                                  41.05        36.62 -10.8%
> ClamAV/clamscan                                35.32        30.90 -12.5%
> kimwitu++/kc                                   30.85        26.98 -12.6%
> SPASS/SPASS                                    33.81        27.72 -18.0%
>                            Geomean difference                      -8.1%
> ```
> 
> ```
> ../utils/compare.py off.json on.json -n .*CTMark.* -m compile_time 
> Tests: 2095
> In filter names: 2085 (filtered out)
> Remaining: 10
> Metric: compile_time
> 
> Program                                       compile_time             
>                                               off          on     diff 
> 7zip/7zip-benchmark                           140.41       146.32  4.2%
> sqlite3/sqlite3                                16.76        16.77  0.1%
> lencod/lencod                                  37.74        37.60 -0.4%
> Bullet/bullet                                  66.20        65.89 -0.5%
> kimwitu++/kc                                   27.76        27.47 -1.0%
> tramp3d-v4/tramp3d-v4                          42.63        42.00 -1.5%
> SPASS/SPASS                                    29.88        28.95 -3.1%
> consumer-typeset/consumer-typeset              21.16        20.35 -3.8%
> ClamAV/clamscan                                34.03        31.69 -6.9%
> mafft/pairlocalalign                           16.99        15.57 -8.3%
>                            Geomean difference                     -2.2%
> ```
> 
> ```
> ../utils/compare.py off.json on.json -n .*CTMark.* -m compile_time 
> Tests: 2095
> In filter names: 2085 (filtered out)
> Remaining: 10
> Metric: compile_time
> 
> Program                                       compile_time             
>                                               off          on     diff 
> 7zip/7zip-benchmark                           141.10       151.16  7.1%
> sqlite3/sqlite3                                17.15        17.90  4.4%
> kimwitu++/kc                                   27.51        27.81  1.1%
> lencod/lencod                                  36.38        36.59  0.6%
> SPASS/SPASS                                    28.84        28.69 -0.5%
> ClamAV/clamscan                                32.15        31.38 -2.4%
> tramp3d-v4/tramp3d-v4                          42.93        41.48 -3.4%
> Bullet/bullet                                  67.83        65.40 -3.6%
> mafft/pairlocalalign                           17.19        16.30 -5.2%
> consumer-typeset/consumer-typeset              21.05        19.49 -7.4%
>                            Geomean difference                     -1.0%
> ```
> 
> It is weird that enabling it reduces compile-time...? I don't know if they are noises.

Is this data from 4 different runs?

https://github.com/llvm/llvm-project/pull/178394