This engineer just discovered a hack that gives CPUs 12% faster

AMD Ryzen 5 7600 Faster CPUs From Just Adding Three author Lili Cui <lili.cui@intel.com> committer Cui, Lili <lili.cui@intel.com> Increase the branch misprediction scale for generic tuning from COSTS_N_INSNS (2) to COSTS_N_INSNS (2) + 3. Modern CPUs have deeper pipelines, making branch mispredictions more expensive. Increasing this cost encourages if-conversion, avoiding pipeline stalls from mispredicted branches. This improves 544.nab r (-02) by 12.7% on GNR and 12.1% on Zmver5 with single-copy. gcc/ChangeLog: * config/i386/x86-tune-costs.h (generic_cost): Increase branch mispredict scale from COSTS_N_INSNS (2) to COSTS_N_INSNS (2) + 3. gcc/config/i386/x86-tune-costs.h patch | blob | history diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h index cc9de64394e073cfe86b5c3d91b26aeb9434b8fd..bc3bb69434919d855e6a81e40bc9b --- a/gcc/config/i386/x86-tune-costs.h +++ b/gcc/config/i386/x86-tune-costs.h @@ -4274,7 +4274,7 @@ struct processor_costs generic_cost = { "16", 4, 2, - COSTS_N_INSNS (2), + COSTS_N_INSNS (2) + 3, /* Branch mispredict scale. */ }; /* core_cost should produce code tuned for Core family of CPUs. */ GNU Compiler Collection branch.c 1 if (temperature > 100) { 2 coolDown(); 3 } else { 4 keepRunning(); 5 } delivery.c 1 int delivery(int total) { 2 int fee; 3 if (total > 499) { 4 fee = 0; 5 } else { 6 fee = 40; 7 } 8 return fee; } pick.c 1 int pick(int a, int b, int x, int y) { 2 int r; 3 if (a < b) { 4 r = x * 3 + 7; 5 } else { 6 r = y * 5 - 2; 7 } 8 return r; } CPUs and fast AI with Vector Neural Network Instructions Hongtao Liu, GCC x86 Backend Vector Maintainer, Compiler Engineering Manager Haochan Jiang, Compiler Engineer Lili Cui, Compiler Engineer Hongyu Wang, Compiler Engineering Manager Lingling Kong, Compiler Engineer Zhang Jun, Compiler Engineer Xu Liwei, Compiler Engineer Hui Lin, Compiler Engineer A one line code change to the GNU Compiler Collection (GCC) for its generic x86 tuning is benefiting modern Intel and AMD processors. Intel software engineer Lili Cui was quite worthwhile increasing the cost of the generic x86 tuning for the cost. On the basis of modern CPUs having decides wrong she that's by updated penalty modern time if is guess calculate with all 50 finally threshold stops risky branchless instead for out CPU lot work