[llvm] [SLP]Fix/improve getSpillCost analysis (PR #129258)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Small increase in compilation time:
> 
> ```
> Program                                       compile_time              
>                                               lhs          rhs     diff 
> FP2017spee...96.specrand_fs/996.specrand_fs      1.68         1.71  2.1%
> FP2017speed/638.imagick_s/638.imagick_s        226.73       231.20  2.0%
> FP2017speed/619.lbm_s/619.lbm_s                  2.83         2.86  1.4%
> FP2017rate/538.imagick_r/538.imagick_r         242.40       245.60  1.3%
> FP2017speed/644.nab_s/644.nab_s                 29.59        29.88  1.0%
> FP2017rate/511.povray_r/511.povray_r           201.27       203.20  1.0%
> FP2017rate/526.blender_r/526.blender_r        1933.89      1948.41  0.8%
> FP2017rate/508.namd_r/508.namd_r                83.35        83.94  0.7%
> INT2017rate/502.gcc_r/502.gcc_r               1224.33      1232.64  0.7%
> INT2017rate/557.xz_r/557.xz_r                   88.05        88.55  0.6%
> FP2017rate/544.nab_r/544.nab_r                  31.05        31.11  0.2%
> INT2017rate/520.omnetpp_r/520.omnetpp_r        632.03       632.91  0.1%
> INT2017rate/541.leela_r/541.leela_r             69.24        69.32  0.1%
> INT2017rat...23.xalancbmk_r/523.xalancbmk_r   1542.00      1542.61  0.0%
> INT2017rat...00.perlbench_r/500.perlbench_r    216.01       216.06  0.0%
> INT2017spe...23.xalancbmk_s/623.xalancbmk_s   1545.39      1541.28 -0.3%
> FP2017rate/510.parest_r/510.parest_r          1851.07      1845.63 -0.3%
> INT2017spe...00.perlbench_s/600.perlbench_s    217.85       217.18 -0.3%
> INT2017speed/605.mcf_s/605.mcf_s                12.58        12.52 -0.5%
> INT2017spe...ed/620.omnetpp_s/620.omnetpp_s    634.10       630.76 -0.5%
> INT2017rat...99.specrand_ir/999.specrand_ir      1.81         1.80 -0.5%
> INT2017rate/525.x264_r/525.x264_r               81.47        80.99 -0.6%
> INT2017speed/641.leela_s/641.leela_s            69.51        69.06 -0.6%
> FP2017rate...97.specrand_fr/997.specrand_fr      1.81         1.80 -0.7%
> INT2017speed/625.x264_s/625.x264_s              82.18        81.55 -0.8%
> INT2017speed/657.xz_s/657.xz_s                  88.77        87.85 -1.0%
> INT2017rate/505.mcf_r/505.mcf_r                 12.51        12.38 -1.0%
> INT2017speed/602.gcc_s/602.gcc_s              1234.66      1218.59 -1.3%
> INT2017rat...31.deepsjeng_r/531.deepsjeng_r     27.21        26.75 -1.7%
> FP2017rate/519.lbm_r/519.lbm_r                   2.92         2.87 -1.8%
> INT2017spe...31.deepsjeng_s/631.deepsjeng_s     27.20        26.67 -2.0%
> INT2017spe...98.specrand_is/998.specrand_is      1.84         1.78 -3.4%
>                            Geomean difference                      -0.2%
> ```

Too much noise. E.g., 502.gcc_r and 602.gcc_s actually should have similar results, since the code is the same.

https://github.com/llvm/llvm-project/pull/129258