Due to the limitations of the GHC runtime, we can't get around the
possibility of out-of-memory errors, but we can prevent individual
bitvectors from being too large for the libgmp-backed bignums.
There is now an architecture-dependent check whenever creating a new
`BV` value in the concrete evaluator to ensure the width does not
exceed the GMP limits. If a width is too large, the evaluation returns
to the REPL much like diving by zero.
Previously it would try to convert the recursive value directly
into a VWord constructor, forcing all bits immediately and causing
non-termination. Now it produces a VSeq constructor instead.
This addresses issue #130.
This "fix" adds two locations to the search path for Cryptol modules so that we can find the prelude on both Windows and Unix platforms. This is not ideal and should be replaced in the next version, but should have no negative impact for now unless people have multiple Cryptol.cry files hanging out in their filesystems.
Previously, if one command was a prefix of another command, that would lead to an ambiguous lookup from the command trie. Now there are two ways to look up: one works like previously, which is good for tab-completion, and the other will only return one result if there is an exact match.
This means we can now have multiple forms of each command, so we can explicitly have short forms like :e -> :edit that won't be changed just based on what names we give to future commands.
This closes#90, and also #94 which was blocked on this due to :exhaust and :edit conflicting.
To mark a failing test as a known failure, you should add a file with
a name like this:
TESTNAME.icry.fails
When the test runs, if it fails, then the contents of this file is displayed.
It is a probably a good idea for the contents to reference the ticket where
the failure was reported.
When the problem is fixed, the `.fails` should be removed. Failing to do
so will result in test failure (i.e., a test that passes but has a `.fails`
file is considered an error).
This technically fixes#103, but not in a tremendously satisfying way. What we would like to do is clean up the progress bars for the :check, and report which input gave rise to the exception.
When the prover is set to "any", Cryptol now checks for all available,
supported provers, submits the problem to all of them in parallel, and
uses the result from whichever finishes first.
While this stops us from leaving the REPL due to negative indicies,
there is an underlying issue we should deal with: absurdly large
indicies. In this case, `list @ (~zero : [64]) means we'll iterate
through 2^64 elements.