Code Optimization Technologies IBM Research Haifa Code Alignment

Code Optimization Technologies IBM Research – Haifa Agenda § Background § Code alignment algorithm –General concepts, code chains –Genetic algorithm § Code alignment for Power 6 –Architecture specifics –Evaluation strategies § Results © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Background § Proper code placement strongly impacts – instruction cache performance – branch prediction – instruction fetch mechanism § Previous works – not many sees code alignment as a code chains placement without reordering, using padding of a certain size – used as a complementary optimization, producing mixed results § We propose a profile-guided generic optimization algorithm, producing stable performance gain © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Code Alignment Algorithm – general concepts § Code chain - is a code sequence which is executed more or less continuously with no significant differences in its instructions frequency § Satisfies one of the following properties: –Terminates with unconditional jump or branch via register –Terminates with a conditional branch whose fallthru is taken infrequently © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Code Alignment Algorithm – working on chains § Aligns each chain by inserting non-executable padding between the chains § Working on chains, not basic blocks, – limits code inflation § Profile allows to focus on frequently executed chains – avoid long run time and code inflation § Tries to determine the best position for each given chain © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Code Chains and Around 0 x 100 Alignment offset of size of 3 instructions Chain 1 Instruction buffer boundary 0 x 120 Gap of size of 4 instructions between the chains 0 x 140 Chain 2 0 x 160 © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Code Alignment Algorithm – filtering alignment options § The algorithm works in phases, in each phase a different measure determines the best alignment alternatives § The initial set of alignment options is defined § This set is filtered in several steps with different filter at each step § These filters, or evaluation strategies, are specific to the architecture and model the performance dependency on the code placement § The strategies are applied based on predefined priorities. The next filter will apply only to results which survived previous filters. The next filter results doesn’t override the previous one’s, but refines it © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Power 6 Pipeline The generic pipeline stages of instruction processing: § Fetch : Instructions are copied from the instruction cache or memory into the fetch buffer. § Decode : Instructions in the fetch buffer are interpreted. § Dispatch : Instructions are sent to the appropriate execution units. § Execute : The operations indicated by the instructions are carried out in the execution units. § Complete : At the end of execution, the result of instructions can be forwarded to other pending instructions while the result awaits write back. § Write Back : The results of execution are written to the architected register, cache or memory in program order, and any exceptions are recognized. © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment for Power 6 In-order architecture, static

Code Optimization Technologies IBM Research – Haifa Alignment for Power 6 – architecture specifics § Fetch buffer contains 8 instructions § Instructions which are not to be executed are discarded § “Good” instructions are delivered for dispatch § Dispatch groups are formed, each cycle one group is executed § A new dispatch group starts on the instruction buffer boundary © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment for Power 6 – evaluation strategies Start from 8 possible alignment options. Filter them by: 1. dispatch groups - minimize the number of dispatch groups formed within the chain, normalized by their execution frequency © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment Evaluation Using Grouping Analysis 0 x 00 offset=0 groups=8 offset=1 groups=8 offset=2 groups=7 offset=3 groups=6 offset=4 groups=7 offset=5 groups=7 offset=6 groups=7 offset=7 groups=6 0 x 20 0 x 40 performing Penalty for offset is the sum of execution grouping counters ofanalysis each created group 0 x 60 © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment for Power 6 – evaluation strategies 2. hot targets - Aligns targets of frequently executed branch instructions, that have high incoming control flow - Best case – hot targets are placed on the beginning of the instruction buffer - Worst case – the first instruction of the hot target is the last instruction of the ibuff - a dispatch group with 1 instruction - this is the only executed instruction of the ibuff © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment Evaluation by Aligned Hot Targets Chain 1 0 x 100 Inserting a gap between the chains to place the hot target on the ibuff boundary 0 x 120 Chain 2 Frequently taken target 0 x 140 0 x 160 © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Alignment for Power 6 – evaluation strategies Other possible strategies: § Branch instructions alignment § Reduce dispatch stalls © 2010 IBM Corporation

Code Optimization Technologies IBM Research – Haifa Results § The algorithm was implemented into IBM FDPR-Pro, a profilebased post-link optimizer § In some cases of extremely bad code alignment up to 40% improvement is achieved § Stable performance gain on SPEC 2006 INT 64 benchmarks, running on AIX 6. 1 on Power 6. Applied on top of standard O 3 FDPR-Pro optimization set and showed up to 5% improvement. © 2010 IBM Corporation