Non-PLT control-point calls

[Pavel-Durov]

Summary

This change optimises control point calls by removing Procedure Linkage Table (PLT) overhead, without introducing regressions in JIT‑compiled execution. It mainly improves interpreter execution time, and the effect is most visible when the JIT is disabled.

Benchmarks with disabled jit:

Datum1: yklua (this change)
Datum0: yklua (main)

confidence level: 99%

 Benchmark               Datum0 (ms)   Datum1 (ms)  Ratio  Summary           
 towers/yklua/600        90160 ± 737   86635 ± 768   0.96  3.91% faster      
 bounce/yklua/1500       90919 ± 809   87616 ± 569   0.96  3.63% faster      
 cd/yklua/250           141950 ± 203  137493 ± 252   0.97  3.14% faster      
 json/yklua/100          57641 ± 108   56055 ±1088   0.97  2.75% faster      
 deltablue/yklua/12000   34073 ± 265   33272 ±  99   0.98  2.35% faster      
 havlak/yklua/1500      198463 ± 548  193872 ±1531   0.98  2.31% faster      
 richards/yklua/100     281299 ± 565  274813 ±1201   0.98  2.31% faster      
 storage/yklua/1000      81921 ± 320   80122 ± 863   0.98  2.20% faster      
 permute/yklua/1000      91389 ±3027   88449 ± 309   0.97  indistinguishable 
 list/yklua/1500         76998 ±1422   74650 ±1694   0.97  indistinguishable 
 sieve/yklua/3000       100572 ±3446   97772 ± 886   0.97  indistinguishable 
 nbody/yklua/250000      72544 ±1456   70638 ±1303   0.97  indistinguishable 
 mandelbrot/yklua/500    41724 ±1827   40882 ± 574   0.98  indistinguishable 
 queens/yklua/1000       63764 ± 496   62811 ±1169   0.99  indistinguishable 

Benchmarks with enabled jit:

Datum1: yklua (this change)
Datum0: yklua (main)

confidence level: 99%

 Benchmark              Datum1 (ms)  Datum0 (ms)  Ratio  Summary           
 towers/yklua/600       11518 ± 536  11120 ± 280   0.97  indistinguishable 
 bounce/yklua/1500      10342 ± 220  10171 ± 497   0.98  indistinguishable 
 json/yklua/100         13929 ± 396  13783 ± 415   0.99  indistinguishable 
 richards/yklua/100     47879 ±2085  47475 ±1519   0.99  indistinguishable 
 sieve/yklua/3000        4849 ± 117   4814 ±  19   0.99  indistinguishable 
 cd/yklua/250           36434 ± 497  36292 ± 274   1.00  indistinguishable 
 mandelbrot/yklua/500    1170 ±   3   1170 ±   1   1.00  indistinguishable 
 havlak/yklua/1500      65648 ±2853  65650 ± 995   1.00  indistinguishable 
 deltablue/yklua/12000  10393 ± 118  10447 ± 177   1.01  indistinguishable 
 nbody/yklua/250000      5571 ±  56   5602 ± 263   1.01  indistinguishable 
 list/yklua/1500         9158 ± 541   9224 ± 530   1.01  indistinguishable 
 storage/yklua/1000     28098 ± 158  28330 ± 398   1.01  indistinguishable 
 queens/yklua/1000       6389 ± 177   6474 ± 400   1.01  indistinguishable 
 permute/yklua/1000      9793 ± 408   9962 ±1154   1.02  indistinguishable 

The change is insignificant when JIT is enabled.

Perf stats

Benchmark	Functions	yklua (main)	yklua-fork (this change)
richards	__ykrt_control_point@plt Runtime Overhead	1.43%	0%
richards	__ykrt_control_point@plt Samples	78	0
havlak	__ykrt_control_point@plt Runtime Overhead	0.80%	0%
havlak	__ykrt_control_point@plt Samples	2950	0
lulpeg	__ykrt_control_point@plt Runtime Overhead	1.47%	0%
lulpeg	__ykrt_control_point@plt Samples	3408	0

Related ykllvm PR ykjit/ykllvm#300

[ltratt]

Does this rely on https://github.com/ykjit/ykllvm/pulls or ... ?

[Pavel-Durov]

Does this rely on https://github.com/ykjit/ykllvm/pulls or ... ?

Yes, it rely on PR ykjit/ykllvm#300 (in the description).

[ltratt]

Please squash.

[Pavel-Durov]

Squashed 5312013