wiki:GhciDebugger

Version 9 (modified by mnislaih, 7 years ago) (diff)

Handling of wrapper constructors

Report of the implementation of the GHCi debugger

During the Summer of 2006 I have been working on this project sponsorized by theGoogle SoC initiative. My mentors were Simon Marlow and David Himmelstrup (lemmih).

It has been a lot of fun, and I've learnt a huge amount of things, but the reader must be warned that I am still a beginner in many aspects, and that my knowledge of ghc is very shallow. So please take my words with a bit of perspective.

The contributions of the project have been mainly two:

  • A closure viewer, capable of showing intermediate computations without forcing them, and without depending on types (and of course that excludes dependency on Show instances)
  • To put the basic breakpoint primitive to use in a system of dynamic breakpoints for ghci.

The closure viewer

The closure viewer functionality is provided by the following function at the GHC module:

obtainTerm :: Session -> Bool -> Id -> IO (Maybe Term)

The term datatype is defined at a module RtClosureInspect in the ghci folder. This datatype represents a partially evaluated Haskell value as an annotated tree:

data Term = Term { ty        :: Type 
                 , dc        :: DataCon 
                 , val       :: HValue 
                 , subTerms  :: [Term] }

          | Prim { ty        :: Type
                 , value     :: String }

          | Suspension { ctype    :: ClosureType
                       , mb_ty    :: Maybe Type
                       , val      :: HValue
                       , bound_to :: Maybe Name   -- Does not belong here, but useful for printing
                       }

A few other comments on this module:

  • It is not meant to be included in the stage1 compiler
  • It is imported by GHC, so in order to avoid introducing more cyclical dependencies I've tried to keep all Session related stuff in the GHC module.

Implementation details

Quoting from Simon Marlow in the ghc-cvs list:

(..)being in GHCi, we have all the compiler's information about the code to hand - including full definitions of data types. So for a given constructor application in the heap, we can print a source-code representation of it

DataCon? recovery

The closure viewer obtains the heap address of a Haskell value, find out the address of its associated info table, and trace back to the DataCon? corresponding to this info table. This is possible because the ghc runtime allocates a static info table for each and every datacon, so all we have to do is extend the linker with a dictionary relating the static info table addresses to a DataCon? name. Moreover, the ghci linker can load interpreted code containing new data or newtype declarations. So the dynamic linker code is extended in the same way. To sum up:

  • linker.c has a new hashtable for datacons.
  • ghci/Linker.hs has been extended in a similar way. The Persistent Link State datatype now includes a datacons environment. At linkExpr and dynLinkBCOs the environment is extended with _any_ new datacons witnessed.
    • Since this scheme makes no distinction between statically and dynamically loaded info tables a lot of redundancy goes into this environment, maybe it's worth to fix this.

Two new primitive ops have been created which allow to obtain the address of a closure info table and to obtain the closure payload (i.e. if it is a value, the arguments of the datacon).

infoPtr# :: a -> Addr#
closurePayload# :: a -> (# Array# b, ByteArr# #)

The use of these primitives is encapsulated in the RtClosureInspect module, which provides:

getClosureType  :: a -> IO ClosureType
getInfoTablePtr :: a -> Ptr StgInfoTable
getClosureData  :: a -> IO Closure

data Closure = Closure { tipe         :: ClosureType 
                       , infoTable    :: StgInfoTable
                       , ptrs         :: Array Int HValue
                       , nonPtrs      :: ByteArray# 
                       }

data ClosureType = Constr 
                 | Fun 
                 | Thunk Int 
                 | ThunkSelector
                 | Blackhole 
                 | AP 
                 | PAP 
                 | Indirection Int 
                 | Other Int
 deriving (Show, Eq)

The implementation of the datacon recovery stuff is scattered around:

Linker.recoverDataCon :: a -> TcM Name
 |- recoverDCInDynEnv :: a -> IO (Maybe Name)
 |- recoverDCInRTS    :: a -> TcM Name
    |- ObjLink.lookupDataCon :: Ptr StgInfoTable -> IO (Maybe String)

First we must make sure that we are dealing with a whnf value (i.e. a Constr), as opposed to a thunk, fun, indirection, etc. This information is retrieved from the very own info table (StgInfoTable? comes with a Storable instance, defined at ByteCodeItbls?). From here on I will use simply constr to refer to a Constr closure.

Once we have the ability to recover the datacon of a constr and thus its (possibly polymorphic) type, we can construct its tree representation. The payload of a closure is an ordered set of pointers and non pointers (words). For a Constr closure, the non pointers correspond to leafs of the tree, primitive unboxed values, the pointers being the so-called subTerms, references to other closures.

Type reconstruction

obtainTerm recursively traverses all the closures that conform a term. Indirections are followed and suspensions are optionally forced. The only problem here is dealing with types. DataCons? can have polymorphic types which we would want to instantiate, so the knowledge of the datacon only is not enough. There are two other sources of type information:

  1. The typechecker, via the Id argument to obtainTerm.
  2. The concrete types of the subterms, if they are sufficiently evaluated.

The process followed to reconstruct the types of a value as much as possible is:

  1. obtain the subTerms of the value recursively calling obtainTerm with the available type info (dataCon plus typechecker), discovering new type info in the process.
  2. refine the type of the value. This is accomplished with a step of unification between (1) and (2) above, and matching the result with the type of the datacon, obtaining the tyvars, which are used to instantiate. This step obtains the most concrete type.
    • Note that tyvars need renaming to avoid collisions.
  3. refine the type of the subterms (inductively) with the reconstructed type.

Compensating Wrapper Constructors

Worker and Wrapper constructors are a potential headache. The arguments list gets extended with:

  • Existential Dictionaries
  • Type equality coercions (?)
  • Type Class dictionaries

In addition, the types of the arguments may change too, but that is not a problem in our case. We always consider the processed types, not the original ones, since what we are going to find in the heap corresponds to the former.

So to recap, all what is being done to compensate is drop the first (n - m) arguments where: n - # arguments of the original constructor m - # arguments of the wrapper constructor, if any, or worker constructor

About handling suspensions in the interactive environment

The interactive ui uses GHC.obtainTerm to implement the :print and :sprint command. The difference is that :print, additionally, binds suspended values. Thus, suspensions inside semievaluated terms are bound by :print to _txx names in the interactive environment, available for the user.

This is done at InteractiveUI.pprintClosure. Whenever the suspensions are not sufficiently typed, tyvars are substituted with the type GHC.Base.Unknown, which has an associated Show instance that instructs the user to seq them to recover the type.

There are two quirks with the current solution:

  • It does not remember previous bindings. Two consecutive uses of :print will generate two separate bindings for the same thing, generating redundancy and potential confusion. But...
  • since type reconstruction (for polymorphic/untyped things) can eventually happen whenever the suspensions are forced, it is necessary to use :print again to obtain a properly typed binding
    • It is a future work to make ghci do this type reconstruction implicitly on the existing, polymorphic bindings. This would be nice for the _txx things, but even nicer for the local bindings in the context of a breakpoint.

Pretty printing of terms

We want to customize the printing of some stuff, such as Integers, Floats, Doubles, Lists, Tuples, Arrays, and so on.

At the RtClosureInspect module there is some infrastructure to build a custom printer, with a basic custom printer that covers the enumerated types.

In InteractiveUI.hs the function pprintClosure takes advantage of this and makes use of a custom printer that uses Show instances if available.

Recovering non-pointers

This happens at RtClosureInspect.extractUnboxed and might potentially break in some architectures.

Breakpoints

breakpoint Implementation

When compiling to bytecodes, breakpoints are desugared to 'fake' jump functions, i.e. they are not defined anywhere, later in the interactive environment we link them to something:

breakpoint => breakpointJump
breakpointCond => breakpointCondJump
breakpointAuto => breakpointAutoJump

The types would be:

breakpointAutoJump, breakpointJump :: 
		    Int				-- Address of a StablePtr containing the Ids
		 -> [()]			-- Local bindings list
		 -> (String, String, Int)	-- Package, Module and site number
		 -> String   	     		-- Location message (filename + srcSpan)
		 -> b -> b   		   
breakpointCond :: Int -> [()] -> (String,String,Int) -> String -> Bool -> b -> b

They get filled with the pointer to the ids in scope, their values, the site, a message, and the wrapped value in the desugarer. Everything served with the right amounts of unsafeCoerce sauce and TyApp? dressing to make the generated Core lint.

The site number is relevant only for 'auto' breakpoints, explained later. For the other two types of breakpoints its value should be 0.

The desugarer monad has been extended with an OccEnv? of Ids to track the bindings in scope. Of course this environment thing is probably too ad-hoc to use it for anything else. The monad also carries a mutable table of breakpoint sites for the current module. This is explained below.

Default HValues for the Jump functions

The dynamic linker has been modified so that it won't panic if one of the jump functions fails to resolve. Now if the dynamic linker fails to find a HValue for a Name, before looking for a static symbol it will ask

DsBreakpoint.lookupBogusBreakpointVal :: Name -> Maybe HValue

which returns a "just return the wrapped thing" if it is one of the Jump names and Nothing otherwise.

This is necessary because a TH function might contain a call to a breakpoint function So if the module it lives in is compiled to bytecodes, the breakpoints will be desugared to 'jumps'. Whenever this code is spliced, the linker will fail to find the jumpfunctions unless there is a default.

Why didn't I address the problem by forbidding breakpoints inside TH code? I couldn't find an easy solution for this, considering the user is free to put a manual breakpoint wherever.

Why did I introduce the default as a special case in the linker?

I considered other options:

  • Running TH splices in an extended link env. This would probably scatter breakpoint related code deep in the typechecker, and is ugly.
  • Making the 'jump' functions real, by giving them equations and types, maybe in the GHC.Exts module. This solution seemed fine but I wasn't sure of how this would interact with dynamic linking of 'jumps'.

A note about bindings in scope in a breakpoint

While I was trying to get the generated core for a breakpoint to lint, I made the design decision of not making available the things bound in a recursive group in the breakpoint context. This includes lets, wheres, and mdo notation. The latter case however is not enforced: I haven't found the time to work it out yet.

Dynamic Breakpoints

The approach followed here has been the well known 'do the simplest thing that could possibly work'. We instrument the code with 'auto' breakpoints at event sites. Currently event sites are code locations where names are bound, and statements:

  • Binding sites (top level, let/where local bindings, case alternatives, lambda abstractions, etc.)
  • do statements (any variant of them)

The instrumentation is done at the desugarer too, which has been extended accordingly. We distinguish between 'auto' breakpoints, those introduced by the desugarer, and 'normal' breakpoints user created by using the breakpoint function directly.

Overhead

The instrumentation scheme potentially introduces overhead at two stages: compile-time and run-time. Compile-time overhead is unnoticeable for general programs, although there are no benchmarks available to sustain this claim. Run-time overhead is much more noticeable. Run-time overhead has been measured informally to range in between 9x and 25x, depending on the code of the program under consideration.

With an always-on breakpoints scenario in mind, we do a number of things to mitigate this overhead in absence of enabled breakpoints. One of these is to allow a ghc-api client to disable auto breakpoints via the ghc-api functions:

enableAutoBreakpoints  :: Session -> IO ()
disableAutoBreakpoints :: Session -> IO ()

GHCi would keep breakpoints disabled until the user defines the first breakpoint, and thus for normal use we could keep the -fdebugging flag enabled always.

The problem is that to make the implementation of disableAutoBreakpoints (enableAutoBreakpoints resp.) effective at all we need to implement it by relinking the breakpointJumpAuto` function to a new "do nothing" lambda (to the user-set bkptHandler resp.).

This would imply a relink, which is quite annoying to a user of GHCi since any top level bindings are lost. This is why this functionality is only a proof of concept and is disabled for now. I wish I had a better understanding of how the dynamic linker and the top level environment in ghci work.

We also try to do some simple breakpoint coalescing.

Breakpoint coalescing

.. implemented, to be documented..

Modifications in the renamer

This section is easy. There are NO modifications in the renamer, other than removing Lemmih's original code for the breakpoint function. All the stuff that we had originally placed here was moved to the desugarer in the final stage of the project.

Modifications to the desugarer

summarize the code instrumentation stuff

Passing the sitelist of a module around

summarize the modifications made to thread the site list of a module from the renamer to the ghc-api

TcGblEnv? is extended with a dictionary of sites and coordinates (TODO switch the coordinate datatype to the ghc-standard SrcLoc?) introduced in the module at the desugarer.

The Opt_Debugging flag

This is activated in command-line via -fdebugging and can be disabled with -fno-debugging. This flag simply enables breakpoint instrumentation in the desugarer.

-fno-debugging is different from -fignore-breakpoints in that user inserted breakpoints will still work.

Interrupting at exceptions

Ideally, a breakpoint that would witness an exception would stop the execution, no more questions. Sadly, it seems impossible to 'witness' an exception. Throw and catch are essentially primitives (throw#, throwio# and catch#), we could install an exception handler at every breakpoint site but that:

  • Would add more overhead
  • Would require serious instrumentation to embed everything in IO, and thus
  • Would alter the evaluation order

So it is not doable via this route.

We could try and use some tricks. For instance, in every 'throw' we spot, we insert a breakpoint based on the condition on this throw. In every 'assert' we do the same. But this would see only user exceptions, missing system exceptions (pattern match failures for instance), asynchronous exceptions and others. Which is not acceptable imho.

I don't know if a satisfactory solution is possible with the current scheme for breakpoints.

The breakpoints api at ghc-api

Once an 'auto' breakpoint, that is a breakpoint inserted by the renamer, is hit, an action is taken. There are hooks to customize this behaviour in the ghc-api. The GHC module provides:

data BkptHandler a = BkptHandler {
     -- | What to do once an enabled breakpoint is found
     handleBreakpoint  :: forall b. Session 
                                  -> [(Id,HValue)]        -- * Local bindings and their id's
                                  -> BkptLocation a    -- * Module and Site # 
                                  ->  String                 -- * A SrcLoc string msg
                                  -> b                         -- * The arg. to the breakpoint fun
                                  -> IO b
     -- | Implementors should return True if the breakpoint is enabled
   , isAutoBkptEnabled :: Session 
                                -> BkptLocation a      -- * Module and Site #
                                -> IO Bool
   }

The Ghci debugger is a client of this API as described below.

The D in Dynamic Breakpoints

In order to implement the 'isAutoBkptEnabled' record, when a breakpoint is hit GHCi must find out whether that site is enabled or not. GHCi thus stores a boolean matrix of enabled breakpoint sites. This scheme is realized in Breakpoints.hs:

data BkptTable a  = BkptTable { 
     breakpoints :: Map.Map a (UArray Int Bool)  -- *An array of breaks, indexed by site number
   , sites       :: Map.Map a [[(SiteNumber, Int)]] -- *A list of lines, each line can have zero or more sites, which are annotated with a column number
   }

Since this structure needs to be accessed every time a breakpoint is hit and is modified extremely few times in comparison, the goal is to have as fast access time as possible. All of the overhead in our debugger is going to be caused by this operation.

It's too bad that I haven't explored alternative designs. (Using bits instead of Bools in the matrix? discard the matrix thing and use an IORef in every breakpoint? some clever trick using the FFI?).

Pending work

Call stack traces. Interruption at unexpected conditions (expections).

Rewrite of the Term pretty printer at RtClosureInspect?.hs Rewrite of the type recovery code Put together all the small todos here