1 / 42

Constraint Programming for Compiler Optimization March 2006

Constraint Programming for Compiler Optimization March 2006. Joint work with: Alexander Golynski Alejandro López-Ortiz Abid Malik Jim McInnes Claude-Guy Quimper John Tromp Kent Wilken. Funding: NSERC IBM Canada. Acknowledgements. Optimization problems in compilers.

franz
Télécharger la présentation

Constraint Programming for Compiler Optimization March 2006

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Constraint Programming for Compiler OptimizationMarch 2006

  2. Joint work with: Alexander Golynski Alejandro López-Ortiz Abid Malik Jim McInnes Claude-Guy Quimper John Tromp Kent Wilken Funding: NSERC IBM Canada Acknowledgements 2

  3. Optimization problems in compilers • Instruction selection • Instruction scheduling • basic-block instruction scheduling • super-block scheduling • software pipelining & loop unrolling • Register allocation • Memory hierarchy optimizations 3

  4. Basic-block instruction scheduling • Schedule basic-block • straight-line sequence of code with single entry, single exit • Multiple-issue pipelined processors • multiple instructions can begin execution each clock cycle • delay or latencybefore results are available • Find minimum length schedule • Classic problem • lots of attention in literature 4

  5. dependency DAG A B 3 3 D C 1 3 E Example: (a + b) + c instructions A r1  a B r2  b C r3  c D r1  r1 + r2 E r1  r1 + r3 5

  6. dependency DAG A B 3 3 D C 1 3 E Single-issue pipelined processor non-optimal schedule A r1  a B r2  b nop nop D r1  r1 + r2 C r3  c nop nop E r1  r1 + r3 6

  7. dependency DAG A B 3 3 D C 1 3 E Single-issue pipelined processor optimal schedule A r1  a B r2  b C r3  c nop D r1  r1 + r2 E r1  r1 + r3 7

  8. dependency DAG issue width is 2 A A B B 1 C 3 3 2 D 3 C D 5 1 3 4 E E Multiple-issue pipelined processor 8

  9. dependency DAG D B A A C B 1 3 3 2 D 3 C 5 6 1 3 4 E Multiple-issue pipelined processor issue width is 1+1 E 9

  10. Production compilers “At the outset, note that basic-block scheduling is an NP-hard problem, even with a very simple formulation of the problem, so we must seek an effective heuristic, rather than exact, approach.” Steven Muchnick, Advanced Compiler Design & Implementation, 1997 10

  11. Single-issue Previous 10-40 instructions ILP (Arya, 1985) CP (Ertl & Krall, 1991) up to 1000 instructions ILP (Wilken et al, 2000) Our work up to 2600 instructions 20 × faster Multiple-issue Previous 10-40 instructions ILP (Chang et al., 1997) DP (Kessler, 1998) up to 1000 instructions B&B (Heffernan et al., 2005) Our work up to 2600 instructions 50-fold improvement Optimal approachesstate-of-the-art 11

  12. Constraint programming methodology • Model problem • specify in terms of constraints on acceptable solutions • define/choose constraint model: variables, domains, constraints • Solve model • define/choose search algorithm • define/choose heuristics 12

  13. Constraint programming methodology • Model problem • specify in terms of constraints on acceptable solutions • define/choose constraint model: variables, domains, constraints • Solve model • define/choose search algorithm • define/choose heuristics 13

  14. dependency DAG A B 3 3 D C 1 3 E Minimal constraint model variables A, B, C, D, E domains {1, …, m} constraints D  A + 3 D  B + 3 E  C + 3 E  D + 1 gcc(A, B, C, D, E, width) 14

  15. Bounds consistency constraint propagation variable A B C D E domain [1, 6] [1, 6] [1, 6] [1, 6] [1, 6]  [1, 2]  [1, 3]  [1, 3]  [1, 2]  [1, 3]  [3, 3]  [4, 6]  [4, 5]  [4, 6]  [5, 6]  [6, 6] constraints D  A + 3 D  B + 3 E  C + 3 E  D + 1 gcc(A, B, C, D, E, 1) 15

  16. Improvements to constraint model • 1. Distance constraints • constraints over nodes which define regions • 2. Predecessor and successor constraints • constraints over nodes with multiple predecessors or multiple successors • 3. Safe pruning constraint • global constraint • 4. Dominance constraints • constraints based on graph isomorphism 16

  17. Improvements to constraint model • 1. Distance constraints • constraints over nodes which define regions • 2. Predecessor and successor constraints • constraints over nodes with multiple predecessors or multiple successors • 3. Safe pruning constraint • global constraint • 4. Dominance constraints • constraints based on graph isomorphism 17

  18. Distance constraints: Regions A pair of nodes i, j define a region in a DAG G if: (i) there is more than one path from i to j, and (ii) not all paths from i to j go through some node k distinct from i and j. i j 18

  19. A 1 1 C B 3 3 D E 1 1 1 F G 3 3 H Distance constraints: Estimate 19

  20. A 1 1 A F j+1 j C B j+2 j+3 j+4 j+5 3 3 5 D E 1 1 1 F G 3 3 H Distance constraints: Estimate 20

  21. A 1 1 E H j+1 j C B j+2 j+3 j+4 j+5 3 3 D E 1 1 1 F 5 G 3 3 H Distance constraints: Estimate 21

  22. A 1 1 A H j+6 j+1 j C B j+2 j+3 j+4 j+5 3 3 D E j+7 j+8 j+9 1 1 1 F G 9 3 3 H Distance constraints: Estimate 22

  23. [1,1] A 1 1 [2,3] [2,3] C B 3 3 [5,6] [5,6] D E 1 1 1 [6,7] [6,7] F G 3 3 [10,10] H Distance constraints: Optimal • Estimate: H  A + 9 • Not optimal: A  1 H  10 propagate latency propagate all-diff 23

  24. [1,1] A 1 1 [2,3] [2,3] C B 3 3 [5,6] [5,6] D E 1 1 1 [6,7] [6,7] F G 3 3 [10,10] H Distance constraints: Optimal • Estimate: H  A + 9 • Not optimal: A  1 H  10 propagate latency propagate all-diff inconsistent • Optimal: H  A + 10 24

  25. Improvements to constraint model • 1. Distance constraints • constraints over nodes which define regions • 2. Predecessor and successor constraints • constraints over nodes with multiple predecessors or multiple successors • 3. Safe pruning constraint • global constraint • 4. Dominance constraints • constraints based on graph isomorphism 25

  26. A 7 1 G B F 1 [5,8] 1 1 D H [6,9] [5,9] [5,9] C 3 3 3 [8,12] [9,12] E 2 2 11 Predecessor constraints [4, ] [ ,14] 26

  27. A [4, ] 7 1 6 5 G B F 1 [5,8] 1 7 8 9 1 H [6,9] [5,9] [5,9] D C 3 3 3 [8,12] [9,12] E 2 2 [ ,14] 11 Predecessor constraints  [9,12] 27

  28. A [4, ] 7 1 9 G B 1 [5,8] 1 10 11 12 1 D [6,9] [5,9] [5,9] C 3 3 3 [8,12]  [9,12] F [9,12] E 2 2 [ ,14] 11 H Predecessor constraints  [12,14] 28

  29. [4, ] 7 A 1 6 1 [5,8] B 1 7 8 9 1 [6,9] [5,9] [5,9] C D E 3 3 3 [8,12]  [9,12] F G [9,12] 2 2  [12,14] [ ,14] 11 H Successor constraints  [4,6] 29

  30. Constraint programming methodology • Model problem • specify in terms of constraints on acceptable solutions • define/choose constraint model: variables, domains, constraints • Solve model • define/choose searchalgorithm • define/choose heuristics 30

  31. Solving instances of the model • Use constraints to establish: • lower bound on length m of optimal schedule • min and max of domains of variables • Backtracking search • branches on min(x), min(x)+1, … • interleave with bounds consistency constraint propagation • fallback: singleton consistency on bounds • If no solution found, increment m and repeat search 31

  32. [1,5] [1,5] A B 3 3 [1,5] [1,5] D C 1 3 [1,5] E Solving instances of the model 1 2 4 5 A B C D E 32

  33. [ ] [ ] A B 3 3 [ ] [ ] D C 1 3 [ ] E Solving instances of the model 1 2 4 5 A B C D E 33

  34. [1,6] [1,6] A B 1 2 5 6 A B 3 3 [1,6] [1,6] C D C D 1 3 [1,6] E E Solving instances of the model 34

  35. [1,2] [1,2] A B 3 3 [5,5] [3,3] D C 1 3 [6,6] E Solving instances of the model 1 2 5 6 A B C D E 35

  36. Improvements to constraint solver • Design special purpose constraint propagators • commonly occurring constraints • significantly improve efficiency • Improved algorithms for bounds consistency • all-diff constraint • gcc constraint 36

  37. Comparing all-diff propagators (prototype) Time (sec.) to solve instruction scheduling problems; model includes latency, distance, and all-diff constraints. DC: Régin, 1994; MT: Mehlhorn & Thiel, 2000; BC: IJCAI-2003 37

  38. Comparing gcc propagators (prototype) Time (sec.) to solve instruction scheduling problems; model includes latency and gcc constraints; width is 2. DC: Régin, 1996; vH: van Hentenryck et al., 1992; BC: CP-2003 38

  39. Putting it all together: Experimental results SPEC 2000 & MediaBench Benchmarks Total of 352,111 basic blocks of size 3 or greater Improved = improved schedule over heuristic scheduler Timed out = not solved within 10 minutes 39

  40. Putting it all together: Experimental results SPEC 2000 & MediaBench Benchmarks For basic blocks with improved schedules 40

  41. Conclusions • CP approach to instruction scheduling • Single-issue processors • 20-times faster than previous best optimal approach • Multiple-issue processors • larger and more difficult problems • 50-fold reduction in number of problems that cannot be solved • Constraint propagators • faster all-diff and gcc constraint propagators • useful in many problems 41

  42. Current and future work: Expand scope of problem • Instruction selection • Instruction scheduling • basic-block instruction scheduling • super-block scheduling • software pipelining & loop unrolling • Register allocation • Memory hierarchy optimizations 42

More Related