Diagnosing potential performance issues using LLVM

U cilju rada je bilo upoznati se s LLVM-om i proučiti performanse programa. Uspoređene su performanse optimiziranih i neoptimiziranih programa. Opisane su najvažnije optimizacije koje primjenjuje prevoditelj Clang. Opisane su metode analiziranja programa. Objema dinamičkim metodama (instrumentacija, uzorkovanje) je analiziran program za pronalaženje duplikata unutar skupa podataka. Pomoću prikladnih alata dijagnosticiran je i otklonjen problematični dio koda. Statičkim analizatorom performansi llvm-mca je analizirana funkcija za zbrajanje dvaju vektora. Jedan od rezultata rada, osim stjecanja uvida u način funkcioniranja alata za analiziranje programa, je stjecanje osnova za rad s LLVM-om. Upoznate su njegove mogućnosti i načini na koji omoguće optimizacije neovisne o arhitekturi procesora.Goal of this paper was to get familiar with LLVM and study the performance of computer programs. Comparisons of optimized and unoptimized computer programs are given. Most important optimizations performed by Clang compiler are described. Methods for program analysis are described. With both of dynamic methods (instrumentation, sampling), program for finding duplicates in dataset was analized. With the use of appropriate tools, problematic piece of code was diagnosed and fixed. Function for adding elements of two vectors was analyzed with static performance analyzer llvm-mca. One result of this paper, besides getting insights into inner workings of tools for program analysis, was learning basics of LLVM. Its abilities and its architecture independent optimizations were studied