1. Automatically Proving the Correctness of Compiler Optimizations Sorin Lerner Todd Millstein Craig Chambers University of Washington

2. Goal: correct compilers The compiler is usually part of the trusted computing base. “But I use gcc, and it works great!”

3. gcc-bugs mailing list c/9525: incorrect code generation on SSE2 intrinsics target/7336: [ARM] With -Os option, gcc incorrectly computes the elimination offset optimization/9325: wrong conversion of constants: (int)(float)(int) (INT_MAX) optimization/6537: For -O (but not -O2 or -O0) incorrect assembly is generated optimization/6891: G++ generates incorrect code when -Os is used optimization/8613: [3.2/3.3/3.4 regression] -O2 optimization generates wrong code target/9732: PPC32: Wrong code with -O2 –fPIC c/8224: Incorrect joining of signed and unsigned division … Searched for “incorrect” and “wrong” in the gcc-bugs mailing list. Some of the results: And this is only for February 2003! On a mature compiler!

4. compiler Source Compiled Prog run! input exp- ected output Testing No correctness guarantees: neither for the compiled prog nor for the compiler DIFF To get benefits, must: run over many inputs compile many test cases output

5. Verify each compilation compiler Source Compiled Prog Semantic DIFF Translation validation [Pnueli et al 98, Necula 00] Credible compilation [Rinard 99] Compiler can still have bugs. Compile time increases. “Semantic Diff” is hard.

6. Proving the whole compiler correct compiler Source Compiled Prog Correctness checker

7. Proving the whole compiler correct compiler Correctness checker Correctness checker Option 1: Prove compiler correct by hand. Proofs are long… And hard. Compilers are proven correct as written on paper. What about the implementation? Proof «¬«¬  $  \ r t  l /. Link?

8. Correctness checker Our Approach Our approach: prove compiler correct automatically. Automatic Theorem Prover compiler

9. This seems really hard! Automatic Theorem Prover Task of proving compiler correct Complexity that an automatic theorem prover can handle. Complexity of proving a compiler correct.

10. Making the problem easier Automatic Theorem Prover Task of proving compiler correct

11. Making the problem easier Automatic Theorem Prover Task of proving optimizer correct Only prove optimizer correct. Trust front-end and code- generator.

12. Making the problem easier Automatic Theorem Prover Write optimizations in Cobalt, a domain-specific language. Task of proving optimizer correct

13. Making the problem easier Automatic Theorem Prover Separate correctness from profitability. Write optimizations in Cobalt, a domain-specific language. Task of proving optimizer correct

14. Making the problem easier Write optimizations in Cobalt, a domain-specific language. Separate correctness from profitability. Factor out the hard and common parts of the proof, and prove them once by hand. Automatic Theorem Prover Task of proving optimizer correct

15. Results Cobalt language –realistic C-like IL –implemented const prop and folding, branch folding, CSE, PRE, DAE, partial DAE, and simple forms of points-to analyses Correctness checker for Cobalt opts –using the Simplify theorem prover Execution engine for Cobalt opts –in the Whirlwind compiler

16. Caveats May not be able to express your opt Cobalt: –no interprocedural optimizations for now. –optimizations that build complicated data structures may be difficult to express. A sound Cobalt optimization may be rejected by the correctness checker. Trusted computing base (TCB) includes: –front-end and code-generator, execution engine, correctness checker, proofs done by hand once

17. Outline Overview Forward optimizations (see paper for backwards) –Example: constant propagation –Strategy for proving forward optimizations sound Profitability heuristics Pure analyses

18. y := 5 x := y REPLACE x := 5 statement y := 5 statements that don’t define y statement x := y Constant Prop (straight-line code)

19. Adding arbitrary control flow y := 5 x := y REPLACE x := 5 statement y := 5 statements that don’t define y statement x := y y := 5 is followed by until transform statement to x := 5 if then

20. Constant prop in statement y := 5 statements that don’t define y is followed by until if then transform statement to x := 5 statement x := y English

21. boolean expressions evaluated at nodes in the CFG stmt(Y := C) X := Y followed by until Cobalt versionEnglish version : mayDef(Y) statement y := 5 statements that don’t define y is followed by until if then transform statement to x := 5 statement x := y Constant prop inCobalt X := C

22. Outline Overview Forward optimizations (see paper for backwards) –Example: constant propagation –Strategy for proving forward optimizations sound Profitability heuristics Pure analyses

23. Proving correctness automatically y := 5 x := yx := 5 y := 5 Witnessing region Invariant: y == 5

24. Constant prop revisited stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C Ask a theorem prover to show: 1.A statement satisfying stmt(Y := C) establishes Y == C 2.A statement satisfying :mayDef(Y) maintains Y == C 3.The statements X := Y and X := C have the same semantics in a program state satisfying Y == C X := C

25. Generalize to any forward optimization Ask a theorem prover to show: 1.A statement satisfying  1 establishes P 2.A statement satisfying  2 maintains P 3.The statements s and s’ have the same semantics in a program state satisfying P We showed by hand once that these conditions imply correctness. 11 22 s followed by until with witness P s’

26. Outline Overview Forward optimizations (see paper for backwards) Profitability heuristics Pure analyses

27. Profitability heuristics Optimization correct ) safe to perform any subset of the matching transformations. So far, all transformations were also profitable. In some cases, many transformations are legal, but only a few are profitable.

28. The two pieces of an optimization  1 followed by  2 until s s’ with witness P filtered through choose Transformation pattern: –defines which transformations are legal. Profitability heuristic: –describes which of the legal transformations to actually perform. –does not affect soundness. –can be written in a language of the user’s choice. This way of factoring an optimization is crucial to our ability to prove optimizations sound automatically.

29. Profitability heuristic example: PRE PRE as code duplication followed by CSE

30. Profitability heuristic example: PRE a :=...; b :=...; if (...) { a :=...; x := a + b; } else {... } x := a + b; Code duplication PRE as code duplication followed by CSE

31. Profitability heuristic example: PRE PRE as code duplication followed by CSE a :=...; b :=...; if (...) { a :=...; x := a + b; } else { } x := x := a + b; Code duplication CSE self-assignment removal a + b; x;

32. Profitability heuristic example: PRE a :=...; b :=...; if (...) { a :=...; x := a + b; } else {... } x := a + b; Legal placements of x := a + b Profitable placement

33. Outline Overview Forward optimizations (see paper for backwards) Profitability heuristics Pure analyses

34. Constant prop revisited (again) stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

35. mayDef in Cobalt stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

36. mayDef in Cobalt Very conservative! Can we do better? stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

37. mayDef in Cobalt Very conservative! Can we do better? stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

38. mayDef in Cobalt stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

39. mayDef in Cobalt mayPntTo is a pure analysis. It computes dataflow info, but performs no transformations. stmt(Y := C) : mayDef(Y) X := Y followed by until with witness Y == C X := C

40. mayPntTo in Cobalt addrNotTaken(X) “no location in the store points to X” decl X s mayPntTo(X,Y), : addrNotTaken(Y) stmt(decl X) followed by : stmt(... := &X) defines with witness

41. Future work Improving expressiveness –interprocedural optimizations –one-to-many and many-to-many transformations Inferring the witness Generate specialized compiler binary from the Cobalt sources.

42. Summary and Conclusion Optimizations written in a domain-specific language can be proven correct automatically. Our correctness checker found several subtle bugs in Cobalt optimizations. A good step towards proving compilers correct automatically.