多核编译器 - 杨全胜.ppt

* Where IPO was a static analysis of your program, PGO is a dynamic analysis. The execution-time characteristics are recorded, and this information is fed into the other optimization phases. Basic block ordering – The compiler uses profile information to place frequently executed blocks in a function close together.? This leads to better I-cache utilization because frequently executed code paths are closer together address-wise. Better register allocation – When allocating registers to hold the values of variables and temporaries, register allocation will give priority to the most frequently executed basic blocks.? As a result, hopefully spills of variables and temporaries to memory (when the register allocator runs out of registers) will be more likely to occur in the infrequently executed basic blocks. Better decision of functions to inline – Inlining incurs code size costs.? By having execution counts of basic blocks (and routines), the compiler can decide to inline heavily executed functions.? There is little value in inlining a routine that is minimally executed (assuming that the profile is representative).? Function ordering – It is similar to basic block ordering; it places heavily executed functions closer together (address-wise).? This enables better paging performance as the utilization of instructions brought in with every page should be higher. Switch-statement optimization – The compiler can pull out high frequency cases and perform an explicit compare for these cases before either doing the tests for the lower frequency cases or performing the indirect jump (which is typically costly on an out-of-order machine with branch prediction). Better vectorization decisions – Vectorization can incur code size cost (extra cleanup loop). ?In addition, it does not make sense to vectorize loops with low trip counts.? Profile information can be used to provide the compiler with heavily executed loops that have high trip counts. * Each execution generates one dyn