7.8. How to Investigate Unproved Checks

One of the most challenging aspects of formal verification is the analysis of failed proofs. If GNATprove fails to prove automatically that a run-time check or an assertion holds, there might be various reasons:

  • [CODE] The check or assertion does not hold, because the code is wrong.
  • [ASSERT] The assertion does not hold, because it is incorrect.
  • [SPEC] The check or assertion cannot be proved, because of some missing assertions about the behavior of the program.
  • [MODEL] The check or assertion is not proved because of current limitations in the model used by GNATprove.
  • [TIMEOUT] The check or assertion is not proved because the prover timeouts.
  • [PROVER] The check or assertion is not proved because the prover is not smart enough.

7.8.1. Investigating Incorrect Code or Assertion

The first step is to check whether the code is incorrect [CODE] or the assertion is incorrect [ASSERT], or both. Since run-time checks and assertions can be executed at run time, one way to increase confidence in the correction of the code and assertions is to test the program on representative inputs. The following GNAT switches can be used:

  • -gnato: enable run-time checking of intermediate overflows
  • -gnat-p: reenable run-time checking even if -gnatp was used to suppress all checks
  • -gnata: enable run-time checking of assertions

7.8.2. Investigating Unprovable Properties

The second step is to consider whether the property is provable [SPEC]. A check or assertion might be unprovable because a necessary annotation is missing:

  • the precondition of the enclosing subprogram might be too weak; or
  • the postcondition of a subprogram called might be too weak; or
  • a loop invariant for an enclosing loop might be too weak; or
  • a loop invariant for a loop before the check or assertion might be too weak.

In particular, GNATprove does not look into subprogram bodies, so all the necessary information for calls should be explicit in the subprogram contracts. A focused manual review of the code and assertions can efficiently diagnose many cases of missing annotations. Even when an assertion is quite large, GNATprove precisely locates the part that it cannot prove, which can help figuring out the problem. It may useful to simplify the code during this investigation, for example by adding a simpler assertion and trying to prove it.

GNATprove provides path information that might help the code review. You can display inside the editor the path on which the proof failed, as described in Running GNATprove from GPS. In some cases, a counterexample is also generated on the path, with values of variables which exhibit the problem (see Understanding Counterexamples). In many cases, this is sufficient to spot a missing assertion.

A property can also be conceptually provable, but the model used by GNATprove can currently not reason about it [MODEL]. (See GNATprove Limitations for a list of the current limitations in GNATprove.) In particular using the following features of the language may yield checks that should be true, but cannot be proved:

  • Floating point arithmetic (although using CodePeer integration may help here)
  • The specific value of dispatching calls when the tag is known

To use CodePeer integration, pass the switch --codepeer=on to GNATprove. In those cases where no prover, including CodePeer, can prove the check, the missing information can usually be added using pragma Assume.

It may be difficult sometimes to distinguish between unprovable properties and prover shortcomings (the next section). The most generally useful action to narrow down the issue to its core is to insert assertions in the code that test whether the property (or part of it) can be proved at some specific point in the program. For example, if a postcondition states a property (P or Q), and the implementation contains many branches and paths, try adding assertions that P holds or Q holds where they are expected to hold. This can help distinguish between the two cases:

  • In the case of an unprovable property, this may point to a specific path in the program, and a specific part of the property, which cause the issue.
  • In the case of a prover shortcoming, this may also help provers to manage to prove both the assertion and the property. Then, it is good practice to keep in the code only those assertions that help getting automatic proof, and to remove other assertions that were inserted during interaction.

7.8.3. Investigating Prover Shortcomings

The last step is to investigate if the prover would find a proof given enough time [TIMEOUT] or if another prover can find a proof [PROVER]. To that end, GNATprove provides switch --level, usable either from the command-line (see Running GNATprove from the Command Line), inside GPS (see Running GNATprove from GPS) or inside GNATbench (see Running GNATprove from GNATbench). The level of 0 is only adequate for simple proofs. In general, one should increase the level of proof (up to level 4) until no more automatic proofs can be obtained.

As described in the section about Running GNATprove from the Command Line, switch --level is equivalent to setting directly various lower level switches like --timeout, --prover, and --proof. Hence, one can also set more powerful (and thus leading to longer proof time) values for the individual switches rather than using the predefined combinations set through --level.

Note that for the above experiments, it is quite convenient to use the SPARK ‣ Prove Line or SPARK ‣ Prove Subprogram menus in GPS, as described in Running GNATprove from GPS and Running GNATprove from GNATbench, to get faster results for the desired line or subprogram.

A current limitation of automatic provers is that they don’t handle floating-point arithmetic very precisely, in particular when there are either a lot of operations, or some non-linear operations (multiplication, division, exponentiation). In that case, it may be profitable to use CodePeer integration, which is activated with the switch --codepeer=on, as CodePeer is both fast and precise for proving bounds of floating-point operations.

Another common limitation of automatic provers is that they don’t handle non-linear arithmetic well. For example, they might fail to prove simple checks involving multiplication, division, modulo or exponentiation.

In that case, a user may either:

In the future, GNATprove may provide a user view of the formula passed to the prover, for advanced users to inspect. This view would express in an Ada-like syntax the actual formula whose proof failed, to make it easier for users to interpret it. This format is yet to be defined.

For advanced users, in particular those who would like to do manual proof, we will provide a description of the format of the proof files generated by GNATprove, so that users can understand the actual files passed to the prover. Each individual file is stored under the sub-directory gnatprove of the project object directory (default is the project directory). The file name follows the convention:



  • file is the name of the Ada source file for the check
  • line is the line where the check appears
  • column is the column
  • check is an identifier for the check
  • num is an optional number and identifies different paths through the program, between the start of the subprogram and the location of the check
  • ext is the extension corresponding to the file format chosen. The format of the file depends on the prover used. For example, files for Alt-Ergo are are in Why3 format, and files for CVC4 are in SMTLIB2 format.

For example, the proof files generated for prover Alt-Ergo for a range check at line 160, column 42, of the file f.adb are stored in:


Corresponding proof files generated for prover CVC4 are stored in:


To be able to inspect these files, you should instruct GNATprove to keep them around by adding the switch -d to GNATprove‘s command line. You can also use the switch -v to get a detailed log of which proof files GNATprove is producing and attempting to prove.

7.8.4. Looking at Machine-Parsable GNATprove Output

GNATprove generates files which contain the results of SPARK analysis in machine-parsable form. These files are located in the gnatprove subdirectory of the project object directory, and have the suffix .spark. The structure of these files exposes internal details such as the exact way some checks are proved, therefore the structure of these files may change. Still, we provide here the structure of these files for convenience.

At various places in these files, we refer to entities. These are Ada entities, either subprograms or packages. Entities are defined by their name and their source location (file and line). In JSON this translates to the following dictionary for entities:

{ "name" : string,
  "file" : string,
  "line" : int }

A .spark file is of this form:

{ "spark" : list spark_result,
  "flow"  : list flow_result,
  "proof" : list proof_result }

Each entry is mapped to a list of entries whose format is described below.

The spark_result entry is simply an entity, with an extra field for spark status, so that the entire dictionary looks like this:

spark_result = { "name"  : string,
                 "file"  : string,
                 "line"  : int,
                 "spark" : string }

Field “spark” takes value in “spec”, “all” or “no” to denote respectively that only the spec is in SPARK, both spec/body are in SPARK (or spec is in SPARK for a package without body), or the spec is not in SPARK.

Entries for proof are of the following form:

proof_result =
  { "file"       : string,
    "line"       : int,
    "col"        : int,
    "suppressed" : string,
    "rule"       : string,
    "severity"   : string,
    "tracefile"  : string,
    "check_tree" : list goal,
    "msg_id"     : int,
    "how_proved" : string,
    "entity"     : entity }
  • (“file”, “line”, “col”) describe the source location of the message.

  • “rule” describes the kind of check.

  • “severity” describes the kind status of the message, possible values used by gnatwhy3 are “info”, “low”, “medium”, “high” and “error”.

  • “tracefile” contains the name of a trace file, if any.

  • “entity” contains the entity dictionary for the entity that this check belongs to.

  • “msg_id” - if present indicates that this entry corresponds to a message issued on the commandline, with the exact same msg_id in brackets: “[#12]”

  • “suppressed” - if present, the message is in fact suppressed by a pragma Annotate, and this field contains the justification message.

  • “how_proved” - if present, indicates how the check has been proved (i.e. which prover). Special values are “interval” and “codepeer”, which designate the special interval analysis, done in the frontend, and the CodePeer analysis, respectively. Both have their own column in the summary table.

  • “check_tree” basically contains a copy of the session tree in JSON format. It’s a tree structure whose nodes are goals, transformations and proof attempts:

    goal = { "transformations" : list trans,
             "pa"              : proof_attempt }
    trans = { [transname : goal] }
    proof_attempt = { [prover : infos] }
    infos = { "time"   : float,
              "steps"  : integer,
              "result" : string }

Flow entries are of the same form as for proof. Differences are in the possible values for “rule”, which can only be the ones for flow messages. Also “how_proved” field is never set.

7.8.5. Understanding Proof Strategies

We now explain in more detail how the provers are run on the logical formula(s) generated for a given check, a.k.a. Verification Conditions or VCs.

  • In per_check mode, a single VC is generated for each check at the source level (e.g. an assertion, run-time check, or postcondition); in some cases two VCs can appear. Before attempting proof, this VC is then split into the conjuncts, that is, the parts that are combined with and or and then. All provers are tried on the VCs obtained in this way until one of them proves the VC or no more provers are left.
  • In per_path mode, a VC is generated not only for each check at the source level, but for each path to the check. For example, for an assertion that appears after an if-statement, at least two VCs will be generated - one for each path trough the if-statement. For each such VC, all provers are attempted. Unproved VCs are then further split into their conjuncts, and proof is again attempted.
  • In progressive mode, first the actions described for per_check are tried. For all unproved VCs, the VC is then split into the paths that lead to the check, like for per_path. Each part is then attempted to be proved independently.