ByteCodeGen.hs

{-# LANGUAGE CPP, MagicHash, RecordWildCards, BangPatterns #-}
{-# LANGUAGE DeriveFunctor #-}
{-# LANGUAGE GeneralizedNewtypeDeriving #-}
{-# OPTIONS_GHC -fprof-auto-top #-}
--
--  (c) The University of Glasgow 2002-2006
--

-- | ByteCodeGen: Generate bytecode from Core
module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where

#include "HsVersions.h"

import GhcPrelude

import ByteCodeInstr
import ByteCodeAsm
import ByteCodeTypes

import GHCi
import GHCi.FFI
import GHCi.RemoteTypes
import BasicTypes
import DynFlags
import Outputable
import Platform
import Name
import MkId
import Id
import ForeignCall
import HscTypes
import CoreUtils
import CoreSyn
import PprCore
import Literal
import PrimOp
import CoreFVs
import Type
import RepType
import Kind            ( isLiftedTypeKind )
import DataCon
import TyCon
import Util
import VarSet
import TysPrim
import ErrUtils
import Unique
import FastString
import Panic
import StgCmmClosure    ( NonVoid(..), fromNonVoid, nonVoidIds )
import StgCmmLayout
import SMRep hiding (WordOff, ByteOff, wordsToBytes)
import Bitmap
import OrdList
import Maybes
import VarEnv

import Data.List
import Foreign
import Control.Monad
import Data.Char

import UniqSupply
import Module
import Control.Arrow ( second )

import Control.Exception
import Data.Array
import Data.ByteString (ByteString)
import Data.Map (Map)
import Data.IntMap (IntMap)
import qualified Data.Map as Map
import qualified Data.IntMap as IntMap
import qualified FiniteMap as Map
import Data.Ord
import GHC.Stack.CCS
import Data.Either ( partitionEithers )

-- -----------------------------------------------------------------------------
-- Generating byte code for a complete module

byteCodeGen :: HscEnv
            -> Module
            -> CoreProgram
            -> [TyCon]
            -> Maybe ModBreaks
            -> IO CompiledByteCode
byteCodeGen hsc_env this_mod binds tycs mb_modBreaks
   = withTiming (pure dflags)
                (text "ByteCodeGen"<+>brackets (ppr this_mod))
                (const ()) $ do
        -- Split top-level binds into strings and others.
        -- See Note [generating code for top-level string literal bindings].
        let (strings, flatBinds) = partitionEithers $ do
                (bndr, rhs) <- flattenBinds binds
                return $ case exprIsTickedString_maybe rhs of
                    Just str -> Left (bndr, str)
                    _ -> Right (bndr, simpleFreeVars rhs)
        stringPtrs <- allocateTopStrings hsc_env strings

        us <- mkSplitUniqSupply 'y'
        (BcM_State{..}, proto_bcos) <-
           runBc hsc_env us this_mod mb_modBreaks (mkVarEnv stringPtrs) $
             mapM schemeTopBind flatBinds

        when (notNull ffis)
             (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")

        dumpIfSet_dyn dflags Opt_D_dump_BCOs
           "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))

        cbc <- assembleBCOs hsc_env proto_bcos tycs (map snd stringPtrs)
          (case modBreaks of
             Nothing -> Nothing
             Just mb -> Just mb{ modBreaks_breakInfo = breakInfo })

        -- Squash space leaks in the CompiledByteCode.  This is really
        -- important, because when loading a set of modules into GHCi
        -- we don't touch the CompiledByteCode until the end when we
        -- do linking.  Forcing out the thunks here reduces space
        -- usage by more than 50% when loading a large number of
        -- modules.
        evaluate (seqCompiledByteCode cbc)

        return cbc

  where dflags = hsc_dflags hsc_env

allocateTopStrings
  :: HscEnv
  -> [(Id, ByteString)]
  -> IO [(Var, RemotePtr ())]
allocateTopStrings hsc_env topStrings = do
  let !(bndrs, strings) = unzip topStrings
  ptrs <- iservCmd hsc_env $ MallocStrings strings
  return $ zip bndrs ptrs

{-
Note [generating code for top-level string literal bindings]

Here is a summary on how the byte code generator deals with top-level string
literals:

1. Top-level string literal bindings are separated from the rest of the module.

2. The strings are allocated via iservCmd, in allocateTopStrings

3. The mapping from binders to allocated strings (topStrings) are maintained in
   BcM and used when generating code for variable references.
-}

-- -----------------------------------------------------------------------------
-- Generating byte code for an expression

-- Returns: the root BCO for this expression
coreExprToBCOs :: HscEnv
               -> Module
               -> CoreExpr
               -> IO UnlinkedBCO
coreExprToBCOs hsc_env this_mod expr
 = withTiming (pure dflags)
              (text "ByteCodeGen"<+>brackets (ppr this_mod))
              (const ()) $ do
      -- create a totally bogus name for the top-level BCO; this
      -- should be harmless, since it's never used for anything
      let invented_name  = mkSystemVarName (mkPseudoUniqueE 0) (fsLit "ExprTopLevel")
          invented_id    = Id.mkLocalId invented_name (panic "invented_id's type")

      -- the uniques are needed to generate fresh variables when we introduce new
      -- let bindings for ticked expressions
      us <- mkSplitUniqSupply 'y'
      (BcM_State _dflags _us _this_mod _final_ctr mallocd _ _ _, proto_bco)
         <- runBc hsc_env us this_mod Nothing emptyVarEnv $
              schemeTopBind (invented_id, simpleFreeVars expr)

      when (notNull mallocd)
           (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")

      dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)

      assembleOneBCO hsc_env proto_bco
  where dflags = hsc_dflags hsc_env

-- The regular freeVars function gives more information than is useful to
-- us here. simpleFreeVars does the impedance matching.
simpleFreeVars :: CoreExpr -> AnnExpr Id DVarSet
simpleFreeVars = go . freeVars
  where
    go :: AnnExpr Id FVAnn -> AnnExpr Id DVarSet
    go (ann, e) = (freeVarsOfAnn ann, go' e)

    go' :: AnnExpr' Id FVAnn -> AnnExpr' Id DVarSet
    go' (AnnVar id)                  = AnnVar id
    go' (AnnLit lit)                 = AnnLit lit
    go' (AnnLam bndr body)           = AnnLam bndr (go body)
    go' (AnnApp fun arg)             = AnnApp (go fun) (go arg)
    go' (AnnCase scrut bndr ty alts) = AnnCase (go scrut) bndr ty (map go_alt alts)
    go' (AnnLet bind body)           = AnnLet (go_bind bind) (go body)
    go' (AnnCast expr (ann, co))     = AnnCast (go expr) (freeVarsOfAnn ann, co)
    go' (AnnTick tick body)          = AnnTick tick (go body)
    go' (AnnType ty)                 = AnnType ty
    go' (AnnCoercion co)             = AnnCoercion co

    go_alt (con, args, expr) = (con, args, go expr)

    go_bind (AnnNonRec bndr rhs) = AnnNonRec bndr (go rhs)
    go_bind (AnnRec pairs)       = AnnRec (map (second go) pairs)

-- -----------------------------------------------------------------------------
-- Compilation schema for the bytecode generator

type BCInstrList = OrdList BCInstr

newtype ByteOff = ByteOff Int
    deriving (Enum, Eq, Integral, Num, Ord, Real)

newtype WordOff = WordOff Int
    deriving (Enum, Eq, Integral, Num, Ord, Real)

wordsToBytes :: DynFlags -> WordOff -> ByteOff
wordsToBytes dflags = fromIntegral . (* wORD_SIZE dflags) . fromIntegral

-- Used when we know we have a whole number of words
bytesToWords :: DynFlags -> ByteOff -> WordOff
bytesToWords dflags (ByteOff bytes) =
    let (q, r) = bytes `quotRem` (wORD_SIZE dflags)
    in if r == 0
           then fromIntegral q
           else panic $ "ByteCodeGen.bytesToWords: bytes=" ++ show bytes

wordSize :: DynFlags -> ByteOff
wordSize dflags = ByteOff (wORD_SIZE dflags)

type Sequel = ByteOff -- back off to this depth before ENTER

type StackDepth = ByteOff

-- | Maps Ids to their stack depth. This allows us to avoid having to mess with
-- it after each push/pop.
type BCEnv = Map Id StackDepth -- To find vars on the stack

{-
ppBCEnv :: BCEnv -> SDoc
ppBCEnv p
   = text "begin-env"
     $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (Map.toList p))))
     $$ text "end-env"
     where
        pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (bcIdArgRep var)
        cmp_snd x y = compare (snd x) (snd y)
-}

-- Create a BCO and do a spot of peephole optimisation on the insns
-- at the same time.
mkProtoBCO
   :: DynFlags
   -> name
   -> BCInstrList
   -> Either  [AnnAlt Id DVarSet] (AnnExpr Id DVarSet)
   -> Int
   -> Word16
   -> [StgWord]
   -> Bool      -- True <=> is a return point, rather than a function
   -> [FFIInfo]
   -> ProtoBCO name
mkProtoBCO dflags nm instrs_ordlist origin arity bitmap_size bitmap is_ret ffis
   = ProtoBCO {
        protoBCOName = nm,
        protoBCOInstrs = maybe_with_stack_check,
        protoBCOBitmap = bitmap,
        protoBCOBitmapSize = bitmap_size,
        protoBCOArity = arity,
        protoBCOExpr = origin,
        protoBCOFFIs = ffis
      }
     where
        -- Overestimate the stack usage (in words) of this BCO,
        -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
        -- stack check.  (The interpreter always does a stack check
        -- for iNTERP_STACK_CHECK_THRESH words at the start of each
        -- BCO anyway, so we only need to add an explicit one in the
        -- (hopefully rare) cases when the (overestimated) stack use
        -- exceeds iNTERP_STACK_CHECK_THRESH.
        maybe_with_stack_check
           | is_ret && stack_usage < fromIntegral (aP_STACK_SPLIM dflags) = peep_d
                -- don't do stack checks at return points,
                -- everything is aggregated up to the top BCO
                -- (which must be a function).
                -- That is, unless the stack usage is >= AP_STACK_SPLIM,
                -- see bug #1466.
           | stack_usage >= fromIntegral iNTERP_STACK_CHECK_THRESH
           = STKCHECK stack_usage : peep_d
           | otherwise
           = peep_d     -- the supposedly common case

        -- We assume that this sum doesn't wrap
        stack_usage = sum (map bciStackUse peep_d)

        -- Merge local pushes
        peep_d = peep (fromOL instrs_ordlist)

        peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
           = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
        peep (PUSH_L off1 : PUSH_L off2 : rest)
           = PUSH_LL off1 (off2-1) : peep rest
        peep (i:rest)
           = i : peep rest
        peep []
           = []

argBits :: DynFlags -> [ArgRep] -> [Bool]
argBits _      [] = []
argBits dflags (rep : args)
  | isFollowableArg rep  = False : argBits dflags args
  | otherwise = take (argRepSizeW dflags rep) (repeat True) ++ argBits dflags args

-- -----------------------------------------------------------------------------
-- schemeTopBind

-- Compile code for the right-hand side of a top-level binding

schemeTopBind :: (Id, AnnExpr Id DVarSet) -> BcM (ProtoBCO Name)
schemeTopBind (id, rhs)
  | Just data_con <- isDataConWorkId_maybe id,
    isNullaryRepDataCon data_con = do
    dflags <- getDynFlags
        -- Special case for the worker of a nullary data con.
        -- It'll look like this:        Nil = /\a -> Nil a
        -- If we feed it into schemeR, we'll get
        --      Nil = Nil
        -- because mkConAppCode treats nullary constructor applications
        -- by just re-using the single top-level definition.  So
        -- for the worker itself, we must allocate it directly.
    -- ioToBc (putStrLn $ "top level BCO")
    emitBc (mkProtoBCO dflags (getName id) (toOL [PACK data_con 0, ENTER])
                       (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})

  | otherwise
  = schemeR [{- No free variables -}] (id, rhs)


-- -----------------------------------------------------------------------------
-- schemeR

-- Compile code for a right-hand side, to give a BCO that,
-- when executed with the free variables and arguments on top of the stack,
-- will return with a pointer to the result on top of the stack, after
-- removing the free variables and arguments.
--
-- Park the resulting BCO in the monad.  Also requires the
-- variable to which this value was bound, so as to give the
-- resulting BCO a name.

schemeR :: [Id]                 -- Free vars of the RHS, ordered as they
                                -- will appear in the thunk.  Empty for
                                -- top-level things, which have no free vars.
        -> (Id, AnnExpr Id DVarSet)
        -> BcM (ProtoBCO Name)
schemeR fvs (nm, rhs)
{-
   | trace (showSDoc (
              (char ' '
               $$ (ppr.filter (not.isTyVar).dVarSetElems.fst) rhs
               $$ pprCoreExpr (deAnnotate rhs)
               $$ char ' '
              ))) False
   = undefined
   | otherwise
-}
   = schemeR_wrk fvs nm rhs (collect rhs)

collect :: AnnExpr Id DVarSet -> ([Var], AnnExpr' Id DVarSet)
collect (_, e) = go [] e
  where
    go xs e | Just e' <- bcView e = go xs e'
    go xs (AnnLam x (_,e))
      | typePrimRep (idType x) `lengthExceeds` 1
      = multiValException
      | otherwise
      = go (x:xs) e
    go xs not_lambda = (reverse xs, not_lambda)

schemeR_wrk
    :: [Id]
    -> Id
    -> AnnExpr Id DVarSet
    -> ([Var], AnnExpr' Var DVarSet)
    -> BcM (ProtoBCO Name)
schemeR_wrk fvs nm original_body (args, body)
   = do
     dflags <- getDynFlags
     let
         all_args  = reverse args ++ fvs
         arity     = length all_args
         -- all_args are the args in reverse order.  We're compiling a function
         -- \fv1..fvn x1..xn -> e
         -- i.e. the fvs come first

         -- Stack arguments always take a whole number of words, we never pack
         -- them unlike constructor fields.
         szsb_args = map (wordsToBytes dflags . idSizeW dflags) all_args
         sum_szsb_args  = sum szsb_args
         p_init    = Map.fromList (zip all_args (mkStackOffsets 0 szsb_args))

         -- make the arg bitmap
         bits = argBits dflags (reverse (map bcIdArgRep all_args))
         bitmap_size = genericLength bits
         bitmap = mkBitmap dflags bits
     body_code <- schemeER_wrk sum_szsb_args p_init body

     emitBc (mkProtoBCO dflags (getName nm) body_code (Right original_body)
                 arity bitmap_size bitmap False{-not alts-})

-- introduce break instructions for ticked expressions
schemeER_wrk :: StackDepth -> BCEnv -> AnnExpr' Id DVarSet -> BcM BCInstrList
schemeER_wrk d p rhs
  | AnnTick (Breakpoint tick_no fvs) (_annot, newRhs) <- rhs
  = do  code <- schemeE d 0 p newRhs
        cc_arr <- getCCArray
        this_mod <- moduleName <$> getCurrentModule
        dflags <- getDynFlags
        let idOffSets = getVarOffSets dflags d p fvs
        let breakInfo = CgBreakInfo
                        { cgb_vars = idOffSets
                        , cgb_resty = exprType (deAnnotate' newRhs)
                        }
        newBreakInfo tick_no breakInfo
        dflags <- getDynFlags
        let cc | interpreterProfiled dflags = cc_arr ! tick_no
               | otherwise = toRemotePtr nullPtr
        let breakInstr = BRK_FUN (fromIntegral tick_no) (getUnique this_mod) cc
        return $ breakInstr `consOL` code
   | otherwise = schemeE d 0 p rhs

getVarOffSets :: DynFlags -> StackDepth -> BCEnv -> [Id] -> [(Id, Word16)]
getVarOffSets dflags depth env = catMaybes . map getOffSet
  where
    getOffSet id = case lookupBCEnv_maybe id env of
        Nothing     -> Nothing
        Just offset ->
            -- michalt: I'm not entirely sure why we need the stack
            -- adjustment by 2 here. I initially thought that there's
            -- something off with getIdValFromApStack (the only user of this
            -- value), but it looks ok to me. My current hypothesis is that
            -- this "adjustment" is needed due to stack manipulation for
            -- BRK_FUN in Interpreter.c In any case, this is used only when
            -- we trigger a breakpoint.
            let !var_depth_ws =
                    trunc16W $ bytesToWords dflags (depth - offset) + 2
            in Just (id, var_depth_ws)

truncIntegral16 :: Integral a => a -> Word16
truncIntegral16 w
    | w > fromIntegral (maxBound :: Word16)
    = panic "stack depth overflow"
    | otherwise
    = fromIntegral w

trunc16B :: ByteOff -> Word16
trunc16B = truncIntegral16

trunc16W :: WordOff -> Word16
trunc16W = truncIntegral16

fvsToEnv :: BCEnv -> DVarSet -> [Id]
-- Takes the free variables of a right-hand side, and
-- delivers an ordered list of the local variables that will
-- be captured in the thunk for the RHS
-- The BCEnv argument tells which variables are in the local
-- environment: these are the ones that should be captured
--
-- The code that constructs the thunk, and the code that executes
-- it, have to agree about this layout
fvsToEnv p fvs = [v | v <- dVarSetElems fvs,
                      isId v,           -- Could be a type variable
                      v `Map.member` p]

-- -----------------------------------------------------------------------------
-- schemeE

returnUnboxedAtom
    :: StackDepth
    -> Sequel
    -> BCEnv
    -> AnnExpr' Id DVarSet
    -> ArgRep
    -> BcM BCInstrList
-- Returning an unlifted value.
-- Heave it on the stack, SLIDE, and RETURN.
returnUnboxedAtom d s p e e_rep = do
    dflags <- getDynFlags
    (push, szb) <- pushAtom d p e
    return (push                                 -- value onto stack
           `appOL`  mkSlideB dflags szb (d - s)  -- clear to sequel
           `snocOL` RETURN_UBX e_rep)            -- go

-- Compile code to apply the given expression to the remaining args
-- on the stack, returning a HNF.
schemeE
    :: StackDepth -> Sequel -> BCEnv -> AnnExpr' Id DVarSet -> BcM BCInstrList
schemeE d s p e
   | Just e' <- bcView e
   = schemeE d s p e'

-- Delegate tail-calls to schemeT.
schemeE d s p e@(AnnApp _ _) = schemeT d s p e

schemeE d s p e@(AnnLit lit)     = returnUnboxedAtom d s p e (typeArgRep (literalType lit))
schemeE d s p e@(AnnCoercion {}) = returnUnboxedAtom d s p e V

schemeE d s p e@(AnnVar v)
    | isUnliftedType (idType v) = returnUnboxedAtom d s p e (bcIdArgRep v)
    | otherwise                 = schemeT d s p e

schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
   | (AnnVar v, args_r_to_l) <- splitApp rhs,
     Just data_con <- isDataConWorkId_maybe v,
     dataConRepArity data_con == length args_r_to_l
   = do -- Special case for a non-recursive let whose RHS is a
        -- saturated constructor application.
        -- Just allocate the constructor and carry on
        alloc_code <- mkConAppCode d s p data_con args_r_to_l
        dflags <- getDynFlags
        let !d2 = d + wordSize dflags
        body_code <- schemeE d2 s (Map.insert x d2 p) body
        return (alloc_code `appOL` body_code)

-- General case for let.  Generates correct, if inefficient, code in
-- all situations.
schemeE d s p (AnnLet binds (_,body)) = do
     dflags <- getDynFlags
     let (xs,rhss) = case binds of AnnNonRec x rhs  -> ([x],[rhs])
                                   AnnRec xs_n_rhss -> unzip xs_n_rhss
         n_binds = genericLength xs

         fvss  = map (fvsToEnv p' . fst) rhss

         -- Sizes of free vars
         size_w = trunc16W . idSizeW dflags
         sizes = map (\rhs_fvs -> sum (map size_w rhs_fvs)) fvss

         -- the arity of each rhs
         arities = map (genericLength . fst . collect) rhss

         -- This p', d' defn is safe because all the items being pushed
         -- are ptrs, so all have size 1 word.  d' and p' reflect the stack
         -- after the closures have been allocated in the heap (but not
         -- filled in), and pointers to them parked on the stack.
         offsets = mkStackOffsets d (genericReplicate n_binds (wordSize dflags))
         p' = Map.insertList (zipE xs offsets) p
         d' = d + wordsToBytes dflags n_binds
         zipE = zipEqual "schemeE"

         -- ToDo: don't build thunks for things with no free variables
         build_thunk
             :: StackDepth
             -> [Id]
             -> Word16
             -> ProtoBCO Name
             -> Word16
             -> Word16
             -> BcM BCInstrList
         build_thunk _ [] size bco off arity
            = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
           where
                mkap | arity == 0 = MKAP
                     | otherwise  = MKPAP
         build_thunk dd (fv:fvs) size bco off arity = do
              (push_code, pushed_szb) <- pushAtom dd p' (AnnVar fv)
              more_push_code <-
                  build_thunk (dd + pushed_szb) fvs size bco off arity
              return (push_code `appOL` more_push_code)

         alloc_code = toOL (zipWith mkAlloc sizes arities)
           where mkAlloc sz 0
                    | is_tick     = ALLOC_AP_NOUPD sz
                    | otherwise   = ALLOC_AP sz
                 mkAlloc sz arity = ALLOC_PAP arity sz

         is_tick = case binds of
                     AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
                     _other -> False

         compile_bind d' fvs x rhs size arity off = do
                bco <- schemeR fvs (x,rhs)
                build_thunk d' fvs size bco off arity

         compile_binds =
            [ compile_bind d' fvs x rhs size arity (trunc16W n)
            | (fvs, x, rhs, size, arity, n) <-
                zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
            ]
     body_code <- schemeE d' s p' body
     thunk_codes <- sequence compile_binds
     return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)

-- Introduce a let binding for a ticked case expression. This rule
-- *should* only fire when the expression was not already let-bound
-- (the code gen for let bindings should take care of that).  Todo: we
-- call exprFreeVars on a deAnnotated expression, this may not be the
-- best way to calculate the free vars but it seemed like the least
-- intrusive thing to do
schemeE d s p exp@(AnnTick (Breakpoint _id _fvs) _rhs)
   | isLiftedTypeKind (typeKind ty)
   = do   id <- newId ty
          -- Todo: is emptyVarSet correct on the next line?
          let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyDVarSet, AnnVar id)
          schemeE d s p letExp

   | otherwise
   = do   -- If the result type is not definitely lifted, then we must generate
          --   let f = \s . tick<n> e
          --   in  f realWorld#
          -- When we stop at the breakpoint, _result will have an unlifted
          -- type and hence won't be bound in the environment, but the
          -- breakpoint will otherwise work fine.
          --
          -- NB (#12007) this /also/ applies for if (ty :: TYPE r), where
          --    r :: RuntimeRep is a variable. This can happen in the
          --    continuations for a pattern-synonym matcher
          --    match = /\(r::RuntimeRep) /\(a::TYPE r).
          --            \(k :: Int -> a) \(v::T).
          --            case v of MkV n -> k n
          -- Here (k n) :: a :: Type r, so we don't know if it's lifted
          -- or not; but that should be fine provided we add that void arg.

          id <- newId (mkVisFunTy realWorldStatePrimTy ty)
          st <- newId realWorldStatePrimTy
          let letExp = AnnLet (AnnNonRec id (fvs, AnnLam st (emptyDVarSet, exp)))
                              (emptyDVarSet, (AnnApp (emptyDVarSet, AnnVar id)
                                                    (emptyDVarSet, AnnVar realWorldPrimId)))
          schemeE d s p letExp

   where
     exp' = deAnnotate' exp
     fvs  = exprFreeVarsDSet exp'
     ty   = exprType exp'

-- ignore other kinds of tick
schemeE d s p (AnnTick _ (_, rhs)) = schemeE d s p rhs

schemeE d s p (AnnCase (_,scrut) _ _ []) = schemeE d s p scrut
        -- no alts: scrut is guaranteed to diverge

schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
   | isUnboxedTupleCon dc -- handles pairs with one void argument (e.g. state token)
        -- Convert
        --      case .... of x { (# V'd-thing, a #) -> ... }
        -- to
        --      case .... of a { DEFAULT -> ... }
        -- because the return convention for both are identical.
        --
        -- Note that it does not matter losing the void-rep thing from the
        -- envt (it won't be bound now) because we never look such things up.
   , Just res <- case (typePrimRep (idType bind1), typePrimRep (idType bind2)) of
                   ([], [_])
                     -> Just $ doCase d s p scrut bind2 [(DEFAULT, [], rhs)] (Just bndr)
                   ([_], [])
                     -> Just $ doCase d s p scrut bind1 [(DEFAULT, [], rhs)] (Just bndr)
                   _ -> Nothing
   = res

schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
   | isUnboxedTupleCon dc
   , typePrimRep (idType bndr) `lengthAtMost` 1 -- handles unit tuples
   = doCase d s p scrut bind1 [(DEFAULT, [], rhs)] (Just bndr)

schemeE d s p (AnnCase scrut bndr _ alt@[(DEFAULT, [], _)])
   | isUnboxedTupleType (idType bndr)
   , Just ty <- case typePrimRep (idType bndr) of
       [_]  -> Just (unwrapType (idType bndr))
       []   -> Just voidPrimTy
       _    -> Nothing
       -- handles any pattern with a single non-void binder; in particular I/O
       -- monad returns (# RealWorld#, a #)
   = doCase d s p scrut (bndr `setIdType` ty) alt (Just bndr)

schemeE d s p (AnnCase scrut bndr _ alts)
   = doCase d s p scrut bndr alts Nothing{-not an unboxed tuple-}

schemeE _ _ _ expr
   = pprPanic "ByteCodeGen.schemeE: unhandled case"
               (pprCoreExpr (deAnnotate' expr))

{-
   Ticked Expressions
   ------------------

  The idea is that the "breakpoint<n,fvs> E" is really just an annotation on
  the code. When we find such a thing, we pull out the useful information,
  and then compile the code as if it was just the expression E.

-}

-- Compile code to do a tail call.  Specifically, push the fn,
-- slide the on-stack app back down to the sequel depth,
-- and enter.  Four cases:
--
-- 0.  (Nasty hack).
--     An application "GHC.Prim.tagToEnum# <type> unboxed-int".
--     The int will be on the stack.  Generate a code sequence
--     to convert it to the relevant constructor, SLIDE and ENTER.
--
-- 1.  The fn denotes a ccall.  Defer to generateCCall.
--
-- 2.  (Another nasty hack).  Spot (# a::V, b #) and treat
--     it simply as  b  -- since the representations are identical
--     (the V takes up zero stack space).  Also, spot
--     (# b #) and treat it as  b.
--
-- 3.  Application of a constructor, by defn saturated.
--     Split the args into ptrs and non-ptrs, and push the nonptrs,
--     then the ptrs, and then do PACK and RETURN.
--
-- 4.  Otherwise, it must be a function call.  Push the args
--     right to left, SLIDE and ENTER.

schemeT :: StackDepth   -- Stack depth
        -> Sequel       -- Sequel depth
        -> BCEnv        -- stack env
        -> AnnExpr' Id DVarSet
        -> BcM BCInstrList

schemeT d s p app

   -- Case 0
   | Just (arg, constr_names) <- maybe_is_tagToEnum_call app
   = implement_tagToId d s p arg constr_names

   -- Case 1
   | Just (CCall ccall_spec) <- isFCallId_maybe fn
   = if isSupportedCConv ccall_spec
      then generateCCall d s p ccall_spec fn args_r_to_l
      else unsupportedCConvException


   -- Case 2: Constructor application
   | Just con <- maybe_saturated_dcon
   , isUnboxedTupleCon con
   = case args_r_to_l of
        [arg1,arg2] | isVAtom arg1 ->
                  unboxedTupleReturn d s p arg2
        [arg1,arg2] | isVAtom arg2 ->
                  unboxedTupleReturn d s p arg1
        _other -> multiValException

   -- Case 3: Ordinary data constructor
   | Just con <- maybe_saturated_dcon
   = do alloc_con <- mkConAppCode d s p con args_r_to_l
        dflags <- getDynFlags
        return (alloc_con         `appOL`
                mkSlideW 1 (bytesToWords dflags $ d - s) `snocOL`
                ENTER)

   -- Case 4: Tail call of function
   | otherwise
   = doTailCall d s p fn args_r_to_l

   where
        -- Extract the args (R->L) and fn
        -- The function will necessarily be a variable,
        -- because we are compiling a tail call
      (AnnVar fn, args_r_to_l) = splitApp app

      -- Only consider this to be a constructor application iff it is
      -- saturated.  Otherwise, we'll call the constructor wrapper.
      n_args = length args_r_to_l
      maybe_saturated_dcon
        = case isDataConWorkId_maybe fn of
                Just con | dataConRepArity con == n_args -> Just con
                _ -> Nothing

-- -----------------------------------------------------------------------------
-- Generate code to build a constructor application,
-- leaving it on top of the stack

mkConAppCode
    :: StackDepth
    -> Sequel
    -> BCEnv
    -> DataCon                  -- The data constructor
    -> [AnnExpr' Id DVarSet]    -- Args, in *reverse* order
    -> BcM BCInstrList
mkConAppCode _ _ _ con []       -- Nullary constructor
  = ASSERT( isNullaryRepDataCon con )
    return (unitOL (PUSH_G (getName (dataConWorkId con))))
        -- Instead of doing a PACK, which would allocate a fresh
        -- copy of this constructor, use the single shared version.

mkConAppCode orig_d _ p con args_r_to_l =
    ASSERT( args_r_to_l `lengthIs` dataConRepArity con ) app_code
  where
    app_code = do
        dflags <- getDynFlags

        -- The args are initially in reverse order, but mkVirtHeapOffsets
        -- expects them to be left-to-right.
        let non_voids =
                [ NonVoid (prim_rep, arg)
                | arg <- reverse args_r_to_l
                , let prim_rep = atomPrimRep arg
                , not (isVoidRep prim_rep)
                ]
            (_, _, args_offsets) =
                mkVirtHeapOffsetsWithPadding dflags StdHeader non_voids

            do_pushery !d (arg : args) = do
                (push, arg_bytes) <- case arg of
                    (Padding l _) -> return $! pushPadding