Make the modular solver the default always, and deprecate the top-down solver #2531
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
It turns out not to be the right solution for general private dependencies and is just complicated. However we keep qualified goals, just much simpler. Now dependencies simply inherit the qualification of their parent goal. This gets us closer to the intended behaviour for the --independent-goals feature, and for the simpler case of private dependencies for setup scripts. When not using --independent-goals, the solver behaves exactly as before (tested by comparing solver logs for a hard hackage goal). When using --independent-goals, now every dep of each independent goal is qualified, so the dependencies are solved completely independently (which is actually too much still).
POption annotates a package choice with a "linked to" field. This commit just introduces the datatype and deals with the immediate fallout, it doesn't actually use the field for anything.
This is implemented as a separate pass so that it can be understood independently of the rest of the solver.
In particular, in the definition of dependencyInconsistencies. One slightly annoying thing is that in order to validate an install plan, we need to know if the goals are to be considered independent. This means we need to pass an additional Bool to a few functions; to limit the number of functions where this is necessary, also recorded whether or not goals are independent as part of the InstallPlan itself.
Since we didn't really have a unit test setup for the solver yet, this introduces some basic tests for solver, as well as tests for independent goals specifically.
This address @23Skidoo's comment #2500 (comment)
I don't know why we we constructed this graph manually here rather than calling `graphFromEdges`; it doesn't really matter except that we will want to change the structure of this graph somewhat once we have more fine-grained dependencies, and then the manual construction becomes a bit more painful; easier to use the standard construction.
This commit does nothing but rearrange the Modular.Dependency module into a number of separate sections, so that's a bit clearer to see what's what. No actual code changes here whatsoever.
The ComponentDeps datatype will give us fine-grained information about the dependencies of a package's components. This commit just introduces the datatype, we don't use it anywhere yet.
The modular solver has its own representation for a package (PInfo). In this commit we modify PInfo to keep track of the different kinds of dependencies. This is a bit intricate because the solver also regards top-level goals as dependencies, but of course those dependencies are not part of any 'component' as such, unlike "real" dependencies. We model this by adding a type parameter to FlaggedDeps and go which indicates whether or not we have component information; crucially, underneath flag choices we _always_ have component information available. Consequently, the modular solver itself will not make use of the ComponentDeps datatype (but only using the Component type, classifying components); we will use ComponentDeps when we translate out of the results from the modular solver into cabal-install's main datatypes. We don't yet _return_ fine-grained dependencies from the solver; this will be the subject of the next commit.
In this commit we modify the _output_ of the modular solver (CP, the modular's solver internal version of ConfiguredPackage) to have fine-grained dependency. This doesn't yet modify the rest of cabal-install, so once we translate from CP to ConfiguredPackage we still lose the distinctions between different kinds of dependencies; this will be the topic of the next commit. In the modular solver (and elsewhere) we use Data.Graph to represent the dependency graph (and the reverse dependency graph). However, now that we have more fine-grained dependencies, we really want an _edge-labeled_ graph, which unfortunately it not available in the `containers` package. Therefore I've written a very simple wrapper around Data.Graph that supports edge labels; we don't need many fancy graph algorithms, and can still use Data.Graph on these edged graphs when we want (by calling them on the underlying unlabeled graph), so adding a dependency on `fgl` does not seem worth it.
The crucial change in this commit is the change to PackageFixedDeps to return a
ComponentDeps structure, rather than a flat list of dependencies, as long with
corresponding changes in ConfiguredPackage and ReadyPackage to accomodate this.
We don't actually take _advantage_ of these more fine-grained dependencies yet;
any use of
depends
is now a use of
CD.flatDeps . depends
but we will :)
Note that I have not updated the top-down solver, so in the output of the
top-down solver we cheat and pretend that all dependencies are library
dependencies.
Although we don't use the new setup dependency component anywhere yet, I've replaced all uses of CD.flatDeps with CD.nonSetupDeps. This means that when we do introduce the setup dependencies, all code in Cabal will still use all dependencies except the setup dependencies, just like now. In other words, using the setup dependencies in some places would be a conscious decision; the default is that we leave the behaviour unchanged.
This patch adds it to the package description types and to the parser. There is a new custom setup section which contains the setup script's dependencies. Also add some sanity checks.
(and, therefore, also to the modular solver's output)
By chosing setup dependencies after regular dependencies we get more opportunities for linking setup dependencies against regular dependencies.
The only problematic thing is that when we call `cabal clean` or `cabal
haddock` (and possibly others), _without_ first having called `configure`, we
attempt to build the setup script without calling the solver at all. This means
that if you do, say,
cabal configure
cabal clean
cabal clean
for a package with a custom setup script that really needs setup dependencies
(for instance, because there are two versions of Cabal in the global package DB
and the setup script needs the _older_ one), then first call to `clean` will
succeed, but the second call will fail because we will try to build the setup
script without the solver and that will fail.
This happened independently in a number of places, which was bad; and was about to get worse with the base 3/4 thing.
Never consider flag choices as independent from their package.
(previously the default was the topdown solver for GHC < 7). Also adds a deprecation warning when the topdown solver is selected.
|
Sorry for taking so long ... |
|
So I'm happy switching now that the new solver works for 6.12 and older. And of course it'll help if ghc ships with multiple base versions again in future. |
dcoutts
added a commit
that referenced
this pull request
May 21, 2015
Make the modular solver the default always, and deprecate the top-down solver
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR is a single commit. It looks big because it depends on all the previous PRs.
This PR makes the modular solver the default solver even for GHC < 7, because we can now deal with the base shim package (PR #2530). To make sure that this is a sensible thing to do, I ran both the top-down solver and the modular solver against every package on Hackage using GHC 6.12.3. Here are some statistics:
leakyandseqaid.Performance-wise, the top-down solver usually took around 11 seconds to solve these goals (and rarely faster than that), while the modular solver was an order of magnitude faster:
I took a look at the cases where the modular solver failed to find a solution. Adding
--reorder-goalsdid not help very much: it only solved an additional 28 cases, leaving a 138 packages where the top-down solver found a solution but the modular solver did not. This also slowed down the modular solver quite a bit:(time in seconds on the x-axis on all these plots). I tried running with
--max-backjumps=-1, with a timeout of a maximum of 5 minutes, and on the packages that I tried this actually still did not enable the modular solver to find a solution within the allotted time.The conclusion from all this is, in my opinion, that it's okay to switch the default now: we find solutions for more packages than we did before, and find them much faster. I guess it's not yet okay to remove the top-down solver completely, for the remaining 138 packages, but we can deprecate the top-down solver and see who complains (which is what this PR does). Meanwhile, we could look at these failing packages to see if we can add any heuristics to the modular solver that would enable it to find solutions. For completeness sake, I added these failures to the Wiki.