Skip to content

Latest commit

 

History

History
313 lines (218 loc) · 15.5 KB

README.md

File metadata and controls

313 lines (218 loc) · 15.5 KB

LiquidHaskell

Hackage Hackage-Deps Build Status

This is the development site of the LiquidHaskell formal verification tool.

If you're a LiquidHaskell user (or just curious), you probably want to go to the documentation website instead.

Contributing

This is an open-source project, and we love getting feedback (and patches)!

Reporting a Bug

If something doesn't work as it should, please consider opening a github issue to let us know. If possible, try to:

  • Try to use a descriptive title;
  • State as clearly as possible what is the problem you are facing;
  • Provide a small Haskell file producing the issue;
  • Write down the expected behaviour vs the actual behaviour;
  • Please, let us know which liquidhaskell version you are using.

Your first Pull Request

We are thrilled to get PRs! Please follow these guidelines, as doing so will increase the chances of having your PR accepted:

  • The main LH repo lives here
  • Please create pull requests against the develop branch.
  • Please be sure to include test cases that illustrate the effect of the PR
    • e.g. show new features that that are supported or how it fixes some previous issue
  • If you're making user-visible changes, please also add documentation

Pull requests don't just have to be about code: documentation can often be improved too!

Ask for Help

If you have further questions or you just need help, you can always reach out on our slack channel, google groups mailing list, GitHub issue tracker, or by emailing Ranjit Jhala, Niki Vazou.

General Development Guide

For those diving into the implementation of LiquidHaskell, here are a few tips:

Running the pluging on individual files

cabal build liquidhaskell
cabal exec ghc -- -fplugin=LiquidHaskell FILE.hs

Building

Cabal

cabal build

Faster recompilation

When changing the liquidhaskell-boot library, sometimes we don't want to rebuild liquidhaskell or liquid-vector when testing the changes. In these cases we can set the environment variable LIQUID_DEV_MODE=true when running cabal to skip rebuilding those packages.

DANGER: Note that this can give an invalid result if the changes to liquidhaskell-boot do require rebuilding other liquid* packages.

How To Run Regression Tests

For documentation on the test-driver executable itself, please refer to the README.md in tests/ or run cabal run tests:test-driver -- --help

You can run all the tests by

$ ./scripts/test/test_plugin.sh

You can run a bunch of particular test-groups instead by

$ ./scripts/test/test_plugin.sh <test-group-name1> <test-group-name2> ...

and you can list all the possible test options with

$ ./scripts/test/test_plugin.sh --help

or get a list of just the test groups, one per line, with

$ ./scripts/test/test_plugin.sh --show-all

To pass in specific parameters, you can invoke cabal directly with

$ cabal build tests:<test-group-name> --ghc-options=-fplugin-opt=LiquidHaskell:--no-termination

For example:

$ cabal build tests:unit-neg --ghc-options=-fplugin-opt=LiquidHaskell:--no-termination

Or your favorite number of threads, depending on cores etc.

You can directly extend and run the tests by modifying the files in

tests/harness/

Parallelism in Tests

Tests run in parallel, unless the flag --measure-timings is specified to test_plugin.sh.

How to create performance comparison charts

When liquidhaskell tests run, we can collect timing information with

$ ./scripts/test/test_plugin.sh --measure-timings

Measures will be collected in .dump-timings files under dist-newstyle directory. These can be converted to json data with

cabal v2-build ghc-timings
cabal v2-exec ghc-timings dist-newstyle

which will produce tmp/*.json files.

Then a csv report can be generated from this json files with

cabal v2-run benchmark-timings -- tmp/*.json --phase LiquidHaskell -o summary.csv

On each line, the report will contain the time taken by each test.

Comparison charts in svg format can be generated by invoking

cabal v2-run plot-performance -- -b path_to_before_summary.csv -a path_to_after_summary.csv -s 50 -f "benchmark"

This will generate three files filtered.svg (a subset of tests with a benchmark prefix, enabled by the -f option), top.svg and bot.svg (top 50 speedups and slowdowns over the entire test set, both enabled by the -s option) in the current directory. The -f and -s options can be used/omitted independently. If both are omitted, a single perf.svg will be produced covering the full input test set. Additionally, their effects can be combined by providing a third -c option (this will produce 2 files filtered-top.svg and filtered-bot.svg instead of 3). An optional key -o can be supplied to specify an output directory for the generated files.

There is also a legacy script scripts/plot-performance/chart_perf.sh that can be used to generate comparison charts in both svg and png formats. It requires gnuplot to run and assumes both files contain the same test set. The following command will produce two files perf.svg and perf.png in the current directory.

$ scripts/plot-performance/chart_perf.sh path_to_before_summary.csv path_to_after_summary.csv

The current formatting is optimized for comparing some subsets of the full test run, typically just the benchmarks alone. If one wishes to save time or is not interested in top speedups/slowdowns, the benchmark subset can be obtained by running

$ scripts/test/test_plugin.sh \
    benchmark-stitch-lh \
    benchmark-bytestring \
    benchmark-vector-algorithms \
    benchmark-cse230 \
    benchmark-esop2013 \
    benchmark-icfp15-pos \
    benchmark-icfp15-neg

Miscelaneous tasks

GHC support policy

LH supports only one version of GHC at any given time. This is because LH depends heavily on the ghc library and there is currently no distinction between public API's and API's internal to GHC. There are currently no release notes for the ghc library and breaking changes happen without notice and without deprecation periods. Supporting only one GHC version saves developer time because it obviates the need for #ifdef's throughout the codebase, or for an compatibility layer that becomes increasingly difficult to write as we attempt to support more GHC versions. Porting to newer GHC versions takes less time, the code is easier to read and there is less code duplication.

Users of older versions of GHC can still use older versions of LH.

The GHC.API module

In order to minimize the effort in porting LH to new releases of GHC, we need a way to abstract over breaking changes in the ghc library, which might change substantially with every major GHC release. This is accomplished by the GHC.API module. The idea is that rather than importing multiple ghc modules, LH developers must import this single module in order to write future-proof code. This is especially important for versions of the compiler greater than 9, where the module hierarchy changed substantially, and using the GHC.API makes it easier to support new versions of GHC when they are released.

Fragile import strategy

import Predicate
import TyCoRep

...

-- This will break if 'isEqPrimPred' is (re)moved or the import hierarchy changes.
foo :: Type -> Bool
foo = isEqPrimPred

Recommended import strategy

import qualified Language.Haskell.Liquid.GHC.API as GHC

...

foo :: GHC.Type -> Bool
foo = GHC.isEqPrimPred -- OK.

GHC Plugin Development Guide

This code commentary describes the current architecture for the GHC Plugin that enables LiquidHaskell to check files as part of the normal compilation process. For the sake of this commentary, we refer to the code provided as part of the release/0.8.10.2 branch, commit 9a2f8284c5fe5b18ed0410e842acd3329a629a6b.

GHC.Interface vs GHC.Plugin

The module GHC.Plugin is the main entrypoint for all the plugin functionalities. Whenever possible, this module is reusing common functionalities from the GHC.Interface, which is the original module used to interface LH with the old executable. Generally speaking, the GHC.Interface module is considered "legacy" and it's rarely what one wants to modify. It will probably be removed at some point.

Plugin architecture

Broadly speaking, the Plugin is organised this way: In the typechecking phase, we typecheck and desugar each module via the GHC API in order to extract the unoptimised core binds that are needed by LH to work correctly. This is due to a tension in the design space; from one side LH needs access to the "raw" core binds (binds where primitives types are not unboxed in the presence of a PRAGMA annotation, for example) but yet the user can specify any arbitrary optimisation settings during compilation and we do not want to betray the principle of least expectation by silently compiling the code with -O0. Practically speaking, this introduces some overhead and is far from ideal, but for now it allows us to iterate quickly. This phase is also responsible for:

  • Extracting the [BareSpec][]s associated to any of the dependent modules;
  • Producing the LiftedSpec for the currently-compiled module;
  • Storing the LiftedSpec into an interface annotation for later retrieval;
  • Checking and verifying the module using LH's existing API.

The reason why we do everything in the typechecking phase is also to allow integrations with tools like ghcide. There are a number of differences between the plugin and the operations performed as part of the GHC.Interface, which we are going to outline in the next section.

Differences with the GHC.Interface

  • The GHC.Interface pre-processes the input files and calls into configureGhcTargets trying to build a dependency graph by discovering dependencies the target files might require. Then, from this list any file in the include directory is filtered out, as well as any module which has a "fresh" .bspec file on disk, to save time during checking. In the GHC.Plugin module though we don't do this and for us, essentially, each input file is considered a target, where we exploit the fact GHC will skip recompilation if unnecessary. This also implies that while the GHC.Interface calls into processTargetModule only for target files, the GHC.Plugin has a single, flat function simply called processModule that essentially does the same as GHC.Interface.processModule and GHC.Interface.processTargetModule fused together.

  • While the GHC.Interface sometimes "assembles" a [BareSpec][] by mappending the commSpec (i.e. comment spec) with the LiftedSpec fetched from disk, if any, the Plugin doesn't do this but rather piggybacks on the SpecFinder (described later) to fetch dependencies' specs.

  • There is a difference in how we process LIQUID pragmas. In particular, for the executable they seems to be accumulated "in bulk" i.e. if we are refining a target module A that depends on B, B seems to inherit whichever flags we were using in the target module A. Conversely, the source plugin is "stateless" when it comes to LIQUID options, i.e. it doesn't have memory of past options, what it counts when compiling a module B is the global options and any option this module defines. The analogy is exactly the same as with GHC language extensions, they have either global scope (i.e. default-extensions in the cabal manifest) or local scope (i.e. {-# LANGUAGE ... #-}).

Finding specs for existing modules

This is all done by a specialised module called the SpecFinder. The main exported function is findRelevantSpecs which, given a list of Modules, tries to retrieve the LiftedSpecs associated with them. Typically this is done by looking into the interface files of the input modules, trying to deserialise any LiftedSpec from the interface file's annotations.

General Development FAQs

A new version of GHC is out. How do I support it?

Typically the first thing you might want to do is to run a "clean" cabal build using the latest compiler and "check the damage". If you are lucky, everything works out of the box, otherwise compilation might fail with an error, typically because some ghc API function has been removed/moved/renamed. The way to fix it is to modify the GHC.API shim module and perform any required change, likely by conditionally compiling some code in a CPP block. For minor changes, it's usually enough to perform small changes, but for more tricky migrations it might be necessary to backport some GHC code, or create some patter synonym to deal with changes in type constructors.

Is there a way to run the testsuite for different versions of GHC?

Currently, no. Only one version of GHC is supported and that is the one that can be tested with ./scripts/test/test_plugin.sh.