CHANGES.md

Version 0.8

New Features

SAW now includes experimental support for verifying Java code using JDK 9 or later. Verifying Java code that only uses primitive data types is known to work well, but code that imports certain classes (e.g., String) is known to suffer from issues documented here.

When verifying Java code, the path to Java can be specified with the new --java-bin-dirs/-b command-line option. Alternatively, if --java-bin-dirs is not set, then SAW searches the PATH to find Java. When the path to Java is known, SAW can automatically add system-related JAR files to the JAR path, which eliminates the need to manually specify these files with -j.

The Crucible-based interface to Java verification is now strictly an improvement over the older code base, with the addition of several features:

• Performance of JVM verification is significantly better, as a result of removing some unnecessary instances of rewriting. This improves performance of LLVM verification, as well.

• The new jvm_static_field_is function allows describing the contents of static variables in method specifications.

• The verification code initializes all JVM classes at the start so that initializers don't run at arbitrary intermediate points and clobber static field values specified in preconditions. This means, however, that any proofs about Java code are under the assumption that all class initializers have run before the method under analysis executes.

Now that the Crucible-based verification infrastructure is sufficiently expressive and performant, we have removed all dependencies on the old jvm-verifier library.

On the LLVM side, SAWScript includes a variety of new functions for writing specification blocks:

• The llvm_struct_type and llvm_packed_struct_type functions each construct an LLVM struct type from a list of other LLVM types. This is not to be confused with the existing llvm_struct function, which takes a string as an argument and returns the corresponding alias type (which is often, but not necessarily, defined as a struct type).

• To avoid confusion, a new llvm_alias function now exists, and llvm_struct is now a synonym for llvm_alias. The llvm_struct function continues to be available for now.

• The llvm_pointer : LLVMType -> LLVMType function allows construction of arbitrary LLVM pointer types.

• Two new functions, llvm_points_to_at_type and llvm_conditional_points_to_at_type, mirror llvm_points_to and llvm_conditional_points_to, but cast the pointer to a different type. This may be useful when reading or writing a prefix of a larger array, for example.

• Support for using ABC as an external process is more complete:

  • SAW can now generate Verilog with multiple outputs (from Term values that have tuple or vector result types, for example).

  • The new commands write_aig_external and write_cnf_external generate AIG and CNF files, respectively, by first writing Verilog and then using the available abc executable to bit-blast to the lower-level representation. Corresponding proof tactics, offline_aig_external and offline_cnf_external also exist.

These changes are in preparation for removing the linked-in copy of ABC in a future release.

The saw-remote-api RPC server and associated Python client now have more complete support for LLVM verification, including:

• More complete points-to declarations, matching what is currently available in SAWScript.

• Support for more provers, including the full range of SBV-based and What4-based provers available in SAWScript.

• Support for ghost variables.

• Support for assuming LLVM contracts directly (rather than the previous behavior which would temporarily assume that failed verifications succeeded to determine whether higher-level verifications would still succeed).

• Support for global variables and initializers.

• Support for null pointers.

• Support for array value literals.

• Support for specifying the value of individual struct fields and array elements.

• Support for specifying the alignment of allocations.

Docker images for SAW are now located on GitHub instead of DockerHub.

Changes

The proof management infrastructure in SAWScript is simpler and more consistent than before. Many of these changes are internal, to make the code less error-prone and easier to maintain in the future. Several are user-facing, though:

• The caseProofResult command now passes a Theorem argument to the first function argument, allowing its use as a rewrite rule, for example.

• A new admit tactic exists, which takes a String argument describing why the user has decided omit proof of the goal. This replaces the assume_unsat and assume_valid tactics, which we now recommend against. They will be officially deprecated in a future release, and removed after that.

• Prover tactics (e.g., yices) now return ProofScript () instead of ProofScript SatResult.

• Simpsets can now contain "permutative" rewrite rules, where a rule is permutative if each side of the equality could represent the same set of terms and therefore the rule could repeatedly apply forever. A term ordering is used to prevent looping when these rules are applied.

Bug Fixes

• Verilog generated from rotation operations is now in correct Verilog syntax.

• Closed issues #9, #25, #39, #41, #54, #55, #69, #81, #90, #92, #124, #136, #144, #149, #152, #159, #271, #285, #323, #353, #377, #382, #431, #446, #631, #652, #739, #740, #861, #901, #908, #924, #930, #951, #962, #971, #985, #991, #993, #994, #995, #996, #1003, #1006, #1009, #1021, #1022, #1023, #1031, #1032, #1050, #1051, #1052, #1055, #1056, #1058, #1061, #1062, #1067, #1073, #1083, #1085, #1090, #1091, #1096, #1099, #1101, #1119, #1122, #1123, #1127, #1128, #1132, #1163, and #1164.

Version 0.7

New Features

SAW can now use the ABC prover as an external process in addition to the linked-in version. This change is in preparation for removing the linked-in version in a future release. This change has several parts:

• The new proof tactics w4_abc_verilog and w4_abc_smtlib2 allow using ABC to discharge proof goals using either Verilog or SMT-Lib2 as the intermediate file format, respectively.
• The new offline_verilog tactic writes a proof goal in the subset of Verilog supported by ABC, which can allow the use of custom ABC commands to discharge it.
• The new w4_offline_smtlib2 writes a proof goal in SMT-Lib2 syntax using What4 instead of SBV.
• The new write_verilog command will write a Term to a file in Verilog syntax from the top level of a script.
• The new write_smtlib2_w4 command will write a Term to a file in SMT-Lib2 syntax from the top level of a script, using What4 instead of SBV.
• The new proof tactics sbv_abc, sbv_boolector, sbv_cvc4, sbv_mathsat, sbv_yices, sbv_z3, sbv_unint_cvc4, sbv_unint_yices, and sbv_unint_z3 are now aliases for the same tactics without the sbv_ prefixes in preparation for making the unprefixed tactics refer to the What4 versions of the same functionality.

Java verification using the Crucible symbolic execution engine is now more flexible and performant.

• The new jvm_array_is command specifies the entire contents of an array together.
• The new jvm_field_is command allows disambiguation by type for fields of the same name but different types.
• JVM method names can be disambiguated using type descriptors.
• JVM constructors can be referred to by the name <init>.
• Error messages in JVM verification are significantly more helpful.

These changes, and various internal performance improvements, mean that the ECDSA verification example included with SAW now runs in around 5 minutes on a modern laptop.

New features exist to summarize any verifications performed with SAW. The summarize_verification command will print a list of all functions or methods verified or assumed, and all lemmas proved with commands such as prove. These summaries include the status (verified or assumed) of each item along with the provers used to complete the proof. The -s <filename> flag will instruct SAW to automatically write a summary to the given file, and the -f <format> flag will instruct SAW to use either a human-readable (pretty) format or JSON (json) for consumption by other tools.

An experimental RPC server for SAW now exists, which provides an alternative to SAWScript for controlling SAW execution. A client library for interacting with the server from Python exists here.

Verification of x86 code called from LLVM is now more flexible. The add_x86_preserved_reg command can add a specific register to the set that is assumed to be callee-saved, and path satisfiability checking is now performed when passing True as the fifth argument to crucible_llvm_verify_x86.

The new cryptol_add_path command adds a directory to the search path used when loading Cryptol modules (and following imports within explicitly-loaded modules).

New, shorter names are available for all LLVM commands starting with the crucible_ prefix. The new names use the llvm_ prefix instead. The top-level function crucible_llvm_verify is now llvm_verify, and crucible_llvm_unsafe_assume_spec is llvm_unsafe_assume_spec. The old names are still supported for now. The in-REPL documentation (:? <command>) gives the new name for each command.

Shorter names are available for some saw-script types: CrucibleSetup is now LLVMSetup, CrucibleMethodSpec is now simply LLVMSpec, and JVMMethodSpec is JVMSpec. The old type names are still supported for backward compatibility.

Bug Fixes

• Catch more exceptions at the REPL (issues #138, #560, #745, and #814).

• Ensure global memory is immutable in x86 verification.

• Fix path sat checking (issues #683 and #723).

• Closed issues #61, #138, #158, #162, #212, #215, #265, #314, #319, #399, #412, #422, #439, #456, #544, #560, #562, #683, #686, #719, #723, #745, #814, #844, #845, #856, #866, #874, #888, #901, #902, #904, #911, #920, #936, #937, #945, and #957.

Version 0.6

New Features

• Adedd experimental support for compositional extraction. Previously, crucible_llvm_extract and similar functions could translate very simple imperative functions into Term models but analysis of more complex programs typically required verification of equivalence to Cryptol specifications. Now, the crucible_llvm_compositional_extract function allows extraction of any function that can be specified using a CrucibleSetup block of the sort used for crucible_llvm_verify. In addition, this extraction can be compositional, preserving the call structure that exists in the original program instead of inlining everything.

• Added experimental support for interactive offline proofs using Coq. The write_coq_term function and offline_coq tactic will export the associated Term to a file in Gallina syntax. This file can be imported into a Coq file that can do arbitrarily complex interactive proofs.

• Added experimental support for arrays of symbolic size. The new crucible_array_alloc function specifies the existence of an allocated array where the size is given by its Term argument. The new crucible_points_to_array_prefix function specifies that the given pointer points to (the prefix of) a symbolic array value of a given symbolic size.

• Improved x86 verification capabilities. Verification scripts for x86 functions can now process functions with mutable globals and function calls (which are currently inlined), and can use proof scripts to discharge proof obligations.

• Added a new llvm_sizeof primitive, which works similarly to the sizeof operator in C.

• Added support for the llvm.fshl.i32 funnel shift intrinsic.

• Added experimental support for verification summaries. These summaries list all verifications performed within a script in a concise way that can track arbitrarily complex proof orchestrations. Future releases will include more information in these summaries, and more textual explanation of the assumptions made during each proof.

Other Changes

• Made small improvements to documentation and error messages throughout.

• Improved the performance of expression hashing (closing #674).

• Updated to include Cryptol 2.9.1 and all the associated changes.

Bug Fixes

• Fixed handling of jumps in x86 verification.

• Fixed matching for points-to and return values for x86.

• Closed issues #140, #195, #225, #226, #387, #538, #542, #543, #556, #635, #640, #642, #674, #691, #693, #698, #699, #700, #701, #702, #704, #706, #707, #709, #720, #721, #730, #742, #744, #763, #768, #773, #783, #785, #789, #782, #803, #807, and #815.

Version 0.5

New Features

• Added experimental support for basic, non-compositional verification of x86 machine code for use in conjunction with LLVM verification.

    crucible_llvm_verify_x86 :
      LLVMModule -> String -> String ->
      [(String, Int)] -> Bool -> CrucibleSetup () ->
      TopLevel CrucibleMethodSpec
  

  The first argument specifies the LLVM module containing the caller. The second and third specify the ELF file name and symbol name of the function to be verifier. The fourth specifies the names and sizes (in bytes) of global variables to initialize, and the fifth whether to perform path satisfiability checking. The last argument is the LLVM specification of the calling context against which to verify the function.

• Added support for using the SMT theory of arrays in the LLVM memory model. In some cases, this can significantly improve performance.

    enable_smt_array_memory_model : TopLevel ()
    disable_smt_array_memory_model : TopLevel ()
  

• Added support for specifying alignment in LLVM allocations. The crucible_alloc_aligned and crucible_alloc_readonly_aligned functions allocate read-write and read-only memory regions, respectively, with the specified alignment (in bytes).

• Added a conditional points-to function, crucible_conditional_points_to, that allows an LLVM function to conditionally modify memory, leaving it in its previous state (potentially uninitialized) when the condition is false.

• Added several new options:

  • New functions enable_what4_hash_consing and disable_what4_hash_consing to enable or disable hash consing to increase sub-formula sharing during symbolic execution.

  • New functions enable_crucible_assert_then_assume and disable_crucible_assert_then_assume to control whether predicates are assumed after asserting them during symbolic execution. The default is now to not assume them, whereas previously they were assumed.

  • New command-line option --no-color to print an ASCII logo without ANSI color or Unicode.

Performance Improvements

• Improved performance of the LLVM memory model.

• Improved performance of bitvector operations during symbolic execution.

• Improved performance of rewriting SAWCore terms.

Bug Fixes

• Fixed SBV interface to fail more gracefully when it cannot find the solver executable.

• Fixed SMT-Lib export negation issues.

Fixes issues #286, #489, #491, #564, #580, #583, #585, #586, #587, #590, #592, #594, #597, #598, #602, #603, #612, #613, #622, #626, #633, #638, #639, and #647.

Version 0.4

• Fixed a long-standing soundness issue (#30) in compositional verification of LLVM programs. Previously, a specification for a function that neglected to mention an effect that the function in fact caused could be successfully verified. When verifying a caller of that function, only the effects mentioned in the specification would be used. The fix for this issue may break some proof scripts: any pointer mentioned using crucible_points_to in the initial state of a specification but not in the final state will be assigned a final value of "invalid", and any subsequent reads from the pointer will fail. To fix this issue, make sure that every specification you use provides a final value for every pointer it touches (which in many cases will be the same as its initial value).

• Added an experimental command, llvm_boilerplate, that emits skeleton function specifications for every function defined in an LLVM module. The additional crucible_llvm_array_size_profile command can be used to refine the results of llvm_boilerplate based on the array sizes used by calls that arise from a call to the named top-level function.

• Added support for using the symbolic execution profiler available in Crucible. The enable_crucible_profiling command causes profiling information to be written to the given directory. This can then be visualized using the rendering code available here.

• Added proof tactics to use Yices (w4_unint_yices) and CVC4 (w4_unint_cvc4) through the What4 backend instead of SBV.

• Modified the messages emitted for failed points-to assertions to be in terms of LLVM values.

• Added support for using the SMT array memory model to reason about the LLVM heap. The enable_smt_array_memory_model command enables it for all future proofs.

• LLVM bitcode format support is improved. Versions 3.5 to 9.0 are known to be mostly well-supported. We consider parsing failures with any version newer than 3.5 to be a bug, so please report them on GitHub.

• New experimental model counting commands sharpSAT and approxmc bind to the external tools of the same name. These were mistakenly listed as included in 0.3.

• Built against Cryptol 2.8.0.

• Improved error messages in general.

• Fixed various additional bugs, including #211, #455, #479, #484, #493, #496, #511, #521, #522, #534, #563

Version 0.3

• Java and LLVM verification has been overhauled to use the new Crucible symbolic execution engine. Highlights include:

  • New crucible_llvm_verify and crucible_llvm_extract commands replace llvm_verify and llvm_extract, with a different structure for specification blocks.

  • LLVM verification tracks undefined behavior more carefully and has a more sophisicated memory model. See the manual for more.

  • New, experimental crucible_jvm_verify and crucible_java_extract commands will eventually replace java_verify and java_extract. For the moment, the former two are enabled with the enable_experimental command and the latter two are enabled with enable_deprecated.

  • More flexible specification language allows convenient description of functions that allocate memory, return arbitrary values, expect explicit aliasing, work with NULL pointers, cast between pointers and integers, or work with opaque pointers.

  • Ghost state is supported in LLVM verification, allowing reasoning about certain complex or unavailable code.

  • Verification of LLVM works for a larger subset of the language, which particularly improves support for C++.

• LLVM bitcode format support is greatly improved. Versions 3.5 to 7.0 are known to be mostly well-supported. We consider parsing failures with any version newer than 3.5 to be a bug, so please report them on GitHub.

• Greatly improved error messages throughout.

• Built against Cryptol 2.7.0.

• New model counting commands sharpSAT and approxmc bind to the external tools of the same name.

• New proof script commands allow multiple goals and related proof tactics. See the manual.

• Can be built with Docker, and will be available on DockerHub.

• Includes an Emacs mode.

Version 0.2-dev

• Released under the 3-clause BSD license

• Major improvements to the Java and LLVM verification infrastructure, as described in more detail here:

  • Major refactoring and polish to java_verify and java_symexec
  • Major refactoring and polish to llvm_verify and llvm_symexec
  • Fixed soundness bug in llvm_verify treatment of heap modifications
  • Fixed soundness bug related to java_assert and llvm_assert
  • Support for branch satisfiability checking to be configured
  • Support for some types of allocation in java_verify, enabled with java_allow_alloc
  • Improved support for LLVM structs (including the llvm_struct type for llvm_verify)
  • Support for non-scalar return values in java_verify and java_symexec
  • Support for using java_ensure_eq on fields of return value
  • Access to safety conditions in java_symexec and llvm_symexec
  • New primitives llvm_assert_eq and java_assert_eq

• Some changes to the SAWScript language:

  • Conditional expressions including the keywords if, then, and else, and the new constants true and false
  • New eval_int and eval_bool functions to expose Cryptol bit vectors and Bit values as Int and Bool values in SAWScript
  • Pattern matching for tuples
  • Improvements to pretty printing, including: set_base and set_ascii commands to control the formatting of values; a show function to convert a value to a string without printing it; and the ability to use print or show instead of llvm_browse_module and java_browse_class
  • New built-in functions for processing lists

• New proof backends:

  • A new rme proof tactic, based on the Reed-Muller Expansion normal form for propositional formulas. This tactic is particularly efficient for dealing with polynomials over Galois fields, as used in AES, for instance.

• Linked against the latest Cryptol code, which includes the following changes since release 2.3.0:

  • An extended prelude with more Haskell-like functions
  • Better, more portable seeding for random
  • Performance improvements for symbolically executing tables of constant values
  • Performance improvements for type checking large constants

• Internal improvements:

  • Simplified Cryptol to SAWCore translation
  • Improved performance of Cryptol to SAWCore translation for recursive functions
  • Updated bitcode parser to support some of the changes in LLVM 3.7
  • Many bug fixes
  • Many code cleanups