1. Runtime Optimization with Specialization Johnathon Jamison CS265 Susan Graham 4-30-2003

2. What is Runtime Code Generation (RTCG)? Dynamic addition of code to the instruction stream Restricted to instructions executed directly by hardware

3. Problems with RTCG Reentrant code Portability (HL languages vs. assembly) Data vs. Code issues –Caches and memory –Standard compilation schema Maintainability and understandability

4. Benefits of RTCG Adaptation to the architecture (cache sizes, various latencies, etc.) JIT compilation No profiling (actual data is available) Literals enable optimizations unknown or impossible at runtime Potentially more compact (for caches)

5. Dynamic Compilation Trade-offs Execution time is linear in run count Choice between lower startup cost, lower incremental cost Unoptimized Static Code Low Optimized Dynamic Code Optimized Static Code High Optimized Dynamic Code Input Execution Time

6. Observation Programmers write the common case –blit routines –image display Applications have repetitious data –simulators –Regexp matching Optimizations –sparse matrices

7. One Tack: Specialization Take a piece of code and replace variables with constants Enables various optimizations –Strength reduction –Constant propagation –etc.... Generate explicitly or implicitly Possibly reuse

8. Example Of Specialization int dot_product(int size, int u[], int v[]) { int res = 0; for (i = 0; i < size; i++) { res = res + u[i] * v[i]; } return res; } Suppose size == 5, u == {14,0,38,12,1}

9. Example Of Specialization int dot_product_1(int v[]) { int res = 0; for (i = 0; i < 5 ; i++) { res = res + {14,0,38,12,1}[i] * v[i]; } return res; } Substitute in the values

10. Example Of Specialization int dot_product_1(int v[]) { int res = 0; res = res + 14 * v[0]; res = res + 0 * v[1]; res = res + 38 * v[2]; res = res + 12 * v[3]; res = res + 1 * v[4]; return res; } Unroll the loop

11. Example Of Specialization int dot_product_1(int v[]) { int res; res = 14 * v[0]; res = res + 38 * v[2]; res = res + 12 * v[3]; res = res + v[4]; return res; } Eliminate unneeded code

12. DyC make_static annotation indicates which variables to specialize with respect to @ annotation indicates static loads (a reload is not needed) int dot_product(int size, int u[], int v[]) { make_static(size, u); int res = 0; for (i = 0; i < size; i++) { res = res + u@[i] * v[i]; } return res; }

13. DyC Specializer Each region has a runtime specializer Setup computations are run The values are plugged into holes in code templates The resultant code is optimized The result is translated to machine code and run

14. DyC Optimizations Polyvariant specialization Internal dynamic-to-static promotions Unchecked dispatching Complete loop unrolling Polyvariant division and conditional specialization Static loads and calls Strength reduction, zero and copy propagation, and dead-assignment elimination (precomputed!)

15. DyC Annotations Runtime constants and constant functions Specialization/division should be mono- /poly-variant Disable/enable internal promotions Compile eagerly/lazily downstream of branches Code caching style at merges/promotions Interprocedural specialization

16. Manual Annotation Profile to find potential gains Concentrate on areas with high execution times If unobvious, log values of parameters to find runtime constants Trial and error loop unrolling

17. Applications

18. Optimizations used

19. Break Even Points

20. Performance

21. Speedup without a given feature

22. Calpa A system that automatically generates DyC annotations Profiles code, collecting statistics Analyses results Annotates code Basically, automates what previously was done manually

23. Calpa, Step 1 Instrumentation tool instruments the original binary Executed on representative input Generates summarized value and frequency data Fed into next step

24. The Instrumenter Three types of information collected –Basic block execution frequencies –Variable definitions –Variable uses Points-to info for invalidation of constants necessary for safety uses stored as value/occurrence pairs, with procedure invocation noted, for groups of related values in a procedure

25. Profiling Data

26. Profiling Seconds to hours Naive profiling was sufficient for their purposes, and so left unoptimized Another paper describes more efficient profile gathering

27. Calpa, Step 2 Annotation tool searches possible space of annotations Selects annotations and creates annotated program Passed to DyC, which compiles the program Calpa == policy, DyC == mechanism

28. Canadate Static Variable (CSV) Sets A CSV set is the set of CSVs that make an instruction static Propagate if exactly one definition exists

29. CSV Sets Example i = 0{} L1:if i >= size goto L2{i, size} uelem = u[i]{i, u[]} velem = v[i]{i, v[]} t = uelem * velem{i, u[], v[]} sum =sum + t{i, sum, u[], v[]} i = i + 1{i} goto L1{} L2:

30. Candidate Division (CD) Sets A CD is a set of CSVs Set of static instructions in a CD are those instructions whose CSV sets are subsets of the CD The CD Set is all CDs produced from some combination of the CSV sets No need to consider other CDs (21 out of 32)

31. CD Sets Example {}{i}{i, size}{i, u[]}{i, v[]}{i, u[], v[]}{i, sum, u[], v[]} {i, size, u[]} {i, size, v[]} {i, size, u[], v[]} {i, size, sum, u[], v[]}

32. Search of CD Space The CDs are enumerated, starting with the least variable variables As additional CDs are enumerated, the "best" one is kept The search terminates if –All CDs are enumerated –a time quota expires –the improvement over the "best" so far drops below a threshold

33. Cost Model Specialization cost –Basic block size * # of vals –Loop size * # of values of the induction variable (scale for multiway loops) –Total instruction * instruction generation cost Cache cost –Lookup cost –Hash key construction (# of vars * cost per var) –Except if unchecked policy is used Invalidation cost –Sum of execution frequency * invalidate cost for all invalidation points

34. Benefit Model Runs a simplified DyC analysis Assumes whole procedure specialization (overestimating costs) Count number of saved cycles assuming the given CD Only looks at the critical path (a simplifying assumption) A win if saved cycles > cycle cost

35. Calpa is safe Static, unchecked, and eager annotations are selected when profile information hints at this However, these are unsafe Calpa does invalidations at all points when it could upset safety Also makes pessimistic assumptions about external routines It is always safe to avoid these annotations

36. Testing Tested on previously annotated programs The annotation process was much quicker Found all manual annotations Plus more annotations

37. Annotations found All the manual ones, plus two more –Search key in search program –Vector v in dotproduct The unvarying nature of these variables was an artifact of atypical use But getting good profiling input is someone else's research

38. Related Work Value Profiling –Work that builds an efficient system to profile programs with the aim of using the collected information to drive specialization. They do not do value sequence information collection, which Calpa needs. They also do binary instrumentation. Fabius –This system takes functions that are curried and generate code for partially evaluated functions. Thus, the idiom of currying is leveraged to optimize code at runtime. Tick C –`C extends C with a few additional constructs that allow explicit manual runtime code compilation. You specify exactly what code that you wish to have compiled in C-like fragments. In the spirit of runtime trade-offs, code generation can be in one of two forms, one quick to create, and one more efficient. Tempo –Tempo can either be a source to source compilation, or a source to runtime code generator. It is much more limited in scope than DyC/Calpa. However, it does have an automatic side-effect/alias analysis.