Statistics

Today I tried to run LLVM tests, but I could not do it automatically. I think it has to do with some configuration issues. I run some tests manually, and the results are not good. Firstly, my algorithm does not produces good code (it is producing around 10% more instructions than linear scan); second, it is very slow. The edge analysis is taking most of the time. Take a look into the comparative result generated during the compilation of this program. This coming week I will try to optimize the algorithm. Today I started with some simple coalescing: when necessary to find a register for a definition, if it is a copy instruction, I try to reuse the register in the used virtual, if its live range is dead. I think that the greatest optimization will be instruction folding. I will try to look into this tomorrow.