-----------------------------------------------------------------------------
--
-- Stg to C--: heap management functions
--
-- (c) The University of Glasgow 2004-2006
--
-----------------------------------------------------------------------------

{-# OPTIONS_GHC -Wno-incomplete-uni-patterns #-}

module GHC.StgToCmm.Heap (
        getVirtHp, setVirtHp, setRealHp,
        getHpRelOffset,

        entryHeapCheck, altHeapCheck, noEscapeHeapCheck, altHeapCheckReturnsTo,
        heapStackCheckGen,
        entryHeapCheck',

        mkStaticClosureFields, mkStaticClosure,

        allocDynClosure, allocDynClosureCmm, allocHeapClosure,
        emitSetDynHdr
    ) where

import GHC.Prelude hiding ((<*>))

import GHC.Stg.Syntax
import GHC.Cmm.CLabel
import GHC.StgToCmm.Layout
import GHC.StgToCmm.Utils
import GHC.StgToCmm.Monad
import GHC.StgToCmm.Prof (profDynAlloc, dynProfHdr, staticProfHdr)
import GHC.StgToCmm.Ticky
import GHC.StgToCmm.Closure

import GHC.Cmm.Graph

import GHC.Cmm.Dataflow.Label
import GHC.Runtime.Heap.Layout
import GHC.Cmm.BlockId
import GHC.Cmm
import GHC.Cmm.Utils
import GHC.Types.CostCentre
import GHC.Types.Id.Info( CafInfo(..), mayHaveCafRefs )
import GHC.Types.Id ( Id )
import GHC.Unit
import GHC.Platform
import GHC.Platform.Profile
import GHC.Data.FastString( mkFastString, fsLit )
import GHC.Utils.Panic( sorry )

import Control.Monad (when)
import Data.Maybe (isJust)

-----------------------------------------------------------
--              Initialise dynamic heap objects
-----------------------------------------------------------

allocDynClosure
        :: Maybe Id
        -> CmmInfoTable
        -> LambdaFormInfo
        -> CmmExpr              -- Cost Centre to stick in the object
        -> CmmExpr              -- Cost Centre to blame for this alloc
                                -- (usually the same; sometimes "OVERHEAD")

        -> [(NonVoid StgArg, VirtualHpOffset)]  -- Offsets from start of object
                                                -- ie Info ptr has offset zero.
                                                -- No void args in here
        -> FCode CmmExpr -- returns Hp+n

allocDynClosureCmm
        :: Maybe Id -> CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
        -> [(CmmExpr, ByteOff)]
        -> FCode CmmExpr -- returns Hp+n

-- allocDynClosure allocates the thing in the heap,
-- and modifies the virtual Hp to account for this.
-- The second return value is the graph that sets the value of the
-- returned LocalReg, which should point to the closure after executing
-- the graph.

-- allocDynClosure returns an (Hp+8) CmmExpr, and hence the result is
-- only valid until Hp is changed.  The caller should assign the
-- result to a LocalReg if it is required to remain live.
--
-- The reason we don't assign it to a LocalReg here is that the caller
-- is often about to call regIdInfo, which immediately assigns the
-- result of allocDynClosure to a new temp in order to add the tag.
-- So by not generating a LocalReg here we avoid a common source of
-- new temporaries and save some compile time.  This can be quite
-- significant - see test T4801.


allocDynClosure :: Maybe Id
-> CmmInfoTable
-> LambdaFormInfo
-> CmmExpr
-> CmmExpr
-> [(NonVoid StgArg, Int)]
-> FCode CmmExpr
allocDynClosure Maybe Id
mb_id CmmInfoTable
info_tbl LambdaFormInfo
lf_info CmmExpr
use_cc CmmExpr
_blame_cc [(NonVoid StgArg, Int)]
args_w_offsets = do
  let ([NonVoid StgArg]
args, [Int]
offsets) = forall a b. [(a, b)] -> ([a], [b])
unzip [(NonVoid StgArg, Int)]
args_w_offsets
  [CmmExpr]
cmm_args <- forall (t :: * -> *) (m :: * -> *) a b.
(Traversable t, Monad m) =>
(a -> m b) -> t a -> m (t b)
mapM NonVoid StgArg -> FCode CmmExpr
getArgAmode [NonVoid StgArg]
args     -- No void args
  Maybe Id
-> CmmInfoTable
-> LambdaFormInfo
-> CmmExpr
-> CmmExpr
-> [(CmmExpr, Int)]
-> FCode CmmExpr
allocDynClosureCmm Maybe Id
mb_id CmmInfoTable
info_tbl LambdaFormInfo
lf_info
                     CmmExpr
use_cc CmmExpr
_blame_cc (forall a b. [a] -> [b] -> [(a, b)]
zip [CmmExpr]
cmm_args [Int]
offsets)


allocDynClosureCmm :: Maybe Id
-> CmmInfoTable
-> LambdaFormInfo
-> CmmExpr
-> CmmExpr
-> [(CmmExpr, Int)]
-> FCode CmmExpr
allocDynClosureCmm Maybe Id
mb_id CmmInfoTable
info_tbl LambdaFormInfo
lf_info CmmExpr
use_cc CmmExpr
_blame_cc [(CmmExpr, Int)]
amodes_w_offsets = do
  -- SAY WHAT WE ARE ABOUT TO DO
  let rep :: SMRep
rep = CmmInfoTable -> SMRep
cit_rep CmmInfoTable
info_tbl
  Maybe Id -> SMRep -> LambdaFormInfo -> FCode ()
tickyDynAlloc Maybe Id
mb_id SMRep
rep LambdaFormInfo
lf_info
  let info_ptr :: CmmExpr
info_ptr = CmmLit -> CmmExpr
CmmLit (CLabel -> CmmLit
CmmLabel (CmmInfoTable -> CLabel
cit_lbl CmmInfoTable
info_tbl))
  SMRep -> CmmExpr -> CmmExpr -> [(CmmExpr, Int)] -> FCode CmmExpr
allocHeapClosure SMRep
rep CmmExpr
info_ptr CmmExpr
use_cc [(CmmExpr, Int)]
amodes_w_offsets


-- | Low-level heap object allocation.
allocHeapClosure
  :: SMRep                            -- ^ representation of the object
  -> CmmExpr                          -- ^ info pointer
  -> CmmExpr                          -- ^ cost centre
  -> [(CmmExpr,ByteOff)]              -- ^ payload
  -> FCode CmmExpr                    -- ^ returns the address of the object
allocHeapClosure :: SMRep -> CmmExpr -> CmmExpr -> [(CmmExpr, Int)] -> FCode CmmExpr
allocHeapClosure SMRep
rep CmmExpr
info_ptr CmmExpr
use_cc [(CmmExpr, Int)]
payload = do
  SMRep -> CmmExpr -> FCode ()
profDynAlloc SMRep
rep CmmExpr
use_cc

  Int
virt_hp <- FCode Int
getVirtHp

  -- Find the offset of the info-ptr word
  let info_offset :: Int
info_offset = Int
virt_hp forall a. Num a => a -> a -> a
+ Int
1
            -- info_offset is the VirtualHpOffset of the first
            -- word of the new object
            -- Remember, virtHp points to last allocated word,
            -- ie 1 *before* the info-ptr word of new object.

  CmmExpr
base <- Int -> FCode CmmExpr
getHpRelOffset Int
info_offset
  FastString -> FCode ()
emitComment forall a b. (a -> b) -> a -> b
$ String -> FastString
mkFastString String
"allocHeapClosure"
  CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
emitSetDynHdr CmmExpr
base CmmExpr
info_ptr CmmExpr
use_cc

  -- Fill in the fields
  CmmExpr -> [(CmmExpr, Int)] -> FCode ()
hpStore CmmExpr
base [(CmmExpr, Int)]
payload

  -- Bump the virtual heap pointer
  Profile
profile <- FCode Profile
getProfile
  Int -> FCode ()
setVirtHp (Int
virt_hp forall a. Num a => a -> a -> a
+ Profile -> SMRep -> Int
heapClosureSizeW Profile
profile SMRep
rep)

  forall (m :: * -> *) a. Monad m => a -> m a
return CmmExpr
base


emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
emitSetDynHdr CmmExpr
base CmmExpr
info_ptr CmmExpr
ccs
  = do Profile
profile <- FCode Profile
getProfile
       CmmExpr -> [(CmmExpr, Int)] -> FCode ()
hpStore CmmExpr
base (forall a b. [a] -> [b] -> [(a, b)]
zip (Profile -> [CmmExpr]
header Profile
profile) [Int
0, Profile -> Int
profileWordSizeInBytes Profile
profile ..])
  where
    header :: Profile -> [CmmExpr]
    header :: Profile -> [CmmExpr]
header Profile
profile = [CmmExpr
info_ptr] forall a. [a] -> [a] -> [a]
++ Profile -> CmmExpr -> [CmmExpr]
dynProfHdr Profile
profile CmmExpr
ccs
        -- ToDo: Parallel stuff
        -- No ticky header

-- Store the item (expr,off) in base[off]
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
hpStore :: CmmExpr -> [(CmmExpr, Int)] -> FCode ()
hpStore CmmExpr
base [(CmmExpr, Int)]
vals = do
  Platform
platform <- FCode Platform
getPlatform
  forall (t :: * -> *) (m :: * -> *) a.
(Foldable t, Monad m) =>
t (m a) -> m ()
sequence_ forall a b. (a -> b) -> a -> b
$
    [ CmmExpr -> CmmExpr -> FCode ()
emitStore (Platform -> CmmExpr -> Int -> CmmExpr
cmmOffsetB Platform
platform CmmExpr
base Int
off) CmmExpr
val | (CmmExpr
val,Int
off) <- [(CmmExpr, Int)]
vals ]

-----------------------------------------------------------
--              Layout of static closures
-----------------------------------------------------------

-- Make a static closure, adding on any extra padding needed for CAFs,
-- and adding a static link field if necessary.

mkStaticClosureFields
        :: Profile
        -> CmmInfoTable
        -> CostCentreStack
        -> CafInfo
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
mkStaticClosureFields :: Profile
-> CmmInfoTable
-> CostCentreStack
-> CafInfo
-> [CmmLit]
-> [CmmLit]
mkStaticClosureFields Profile
profile CmmInfoTable
info_tbl CostCentreStack
ccs CafInfo
caf_refs [CmmLit]
payload
  = Profile
-> CLabel
-> CostCentreStack
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
mkStaticClosure Profile
profile CLabel
info_lbl CostCentreStack
ccs [CmmLit]
payload [CmmLit]
padding
        [CmmLit]
static_link_field [CmmLit]
saved_info_field
  where
    platform :: Platform
platform = Profile -> Platform
profilePlatform Profile
profile
    info_lbl :: CLabel
info_lbl = CmmInfoTable -> CLabel
cit_lbl CmmInfoTable
info_tbl

    -- CAFs must have consistent layout, regardless of whether they
    -- are actually updatable or not.  The layout of a CAF is:
    --
    --        3 saved_info
    --        2 static_link
    --        1 indirectee
    --        0 info ptr
    --
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:

    is_caf :: Bool
is_caf = SMRep -> Bool
isThunkRep (CmmInfoTable -> SMRep
cit_rep CmmInfoTable
info_tbl)

    padding :: [CmmLit]
padding
        | Bool
is_caf Bool -> Bool -> Bool
&& forall (t :: * -> *) a. Foldable t => t a -> Bool
null [CmmLit]
payload = [Platform -> Int -> CmmLit
mkIntCLit Platform
platform Int
0]
        | Bool
otherwise = []

    static_link_field :: [CmmLit]
static_link_field
        | Bool
is_caf
        = [Platform -> Int -> CmmLit
mkIntCLit Platform
platform Int
0]
        | Bool -> CmmInfoTable -> Bool
staticClosureNeedsLink (CafInfo -> Bool
mayHaveCafRefs CafInfo
caf_refs) CmmInfoTable
info_tbl
        = [CmmLit
static_link_value]
        | Bool
otherwise
        = []

    saved_info_field :: [CmmLit]
saved_info_field
        | Bool
is_caf     = [Platform -> Int -> CmmLit
mkIntCLit Platform
platform Int
0]
        | Bool
otherwise  = []

        -- For a static constructor which has NoCafRefs, we set the
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
    static_link_value :: CmmLit
static_link_value
        | CafInfo -> Bool
mayHaveCafRefs CafInfo
caf_refs  = Platform -> Int -> CmmLit
mkIntCLit Platform
platform Int
0
        | Bool
otherwise                = Platform -> Int -> CmmLit
mkIntCLit Platform
platform Int
3  -- No CAF refs
                                      -- See Note [STATIC_LINK fields]
                                      -- in rts/sm/Storage.h

mkStaticClosure :: Profile -> CLabel -> CostCentreStack -> [CmmLit]
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
mkStaticClosure :: Profile
-> CLabel
-> CostCentreStack
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
-> [CmmLit]
mkStaticClosure Profile
profile CLabel
info_lbl CostCentreStack
ccs [CmmLit]
payload [CmmLit]
padding [CmmLit]
static_link_field [CmmLit]
saved_info_field
  =  [CLabel -> CmmLit
CmmLabel CLabel
info_lbl]
  forall a. [a] -> [a] -> [a]
++ Profile -> CostCentreStack -> [CmmLit]
staticProfHdr Profile
profile CostCentreStack
ccs
  forall a. [a] -> [a] -> [a]
++ [CmmLit]
payload
  forall a. [a] -> [a] -> [a]
++ [CmmLit]
padding
  forall a. [a] -> [a] -> [a]
++ [CmmLit]
static_link_field
  forall a. [a] -> [a] -> [a]
++ [CmmLit]
saved_info_field

-----------------------------------------------------------
--              Heap overflow checking
-----------------------------------------------------------

{- Note [Heap checks]
   ~~~~~~~~~~~~~~~~~~
Heap checks come in various forms.  We provide the following entry
points to the runtime system, all of which use the native C-- entry
convention.

  * gc() performs garbage collection and returns
    nothing to its caller

  * A series of canned entry points like
        r = gc_1p( r )
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.

  * A series of canned entry points like
        gcfun_2p( f, x, y )
    where f is a function closure of arity 2
    This performs garbage collection, keeping alive the
    three argument ptrs, and then tail-calls f(x,y)

These are used in the following circumstances

* entryHeapCheck: Function entry
    (a) With a canned GC entry sequence
        f( f_clo, x:ptr, y:ptr ) {
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
     it should do no register shuffling

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
          L: HpAlloc = 8
             call gc()  -- Needs an info table
             goto T }

* altHeapCheck: Immediately following an eval
  Started as
        case f x y of r { (p,q) -> rhs }
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...

        L: r = gc_1p( r )
           goto K }

        Here, the info table needed by the call
        to gc_1p should be the *same* as the
        one for the call to f; the C-- optimiser
        spots this sharing opportunity)

   (b) No canned sequence for results of f
       Note second info table
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...

        L: call gc()    -- Extra info table here
           goto K

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
       case branch *not* following eval,
       or let-no-escape
  Exactly the same as the previous case:

        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...

        L: call gc()
           goto K
-}

--------------------------------------------------------------
-- A heap/stack check at a function or thunk entry point.

entryHeapCheck :: ClosureInfo
               -> Maybe LocalReg -- Function (closure environment)
               -> Int            -- Arity -- not same as len args b/c of voids
               -> [LocalReg]     -- Non-void args (empty for thunk)
               -> FCode ()
               -> FCode ()

entryHeapCheck :: ClosureInfo
-> Maybe LocalReg -> Int -> [LocalReg] -> FCode () -> FCode ()
entryHeapCheck ClosureInfo
cl_info Maybe LocalReg
nodeSet Int
arity [LocalReg]
args FCode ()
code = do
  Platform
platform <- FCode Platform
getPlatform
  let
    node :: CmmExpr
node = case Maybe LocalReg
nodeSet of
              Just LocalReg
r  -> CmmReg -> CmmExpr
CmmReg (LocalReg -> CmmReg
CmmLocal LocalReg
r)
              Maybe LocalReg
Nothing -> CmmLit -> CmmExpr
CmmLit (CLabel -> CmmLit
CmmLabel forall a b. (a -> b) -> a -> b
$ Platform -> ClosureInfo -> CLabel
staticClosureLabel Platform
platform ClosureInfo
cl_info)

    is_fastf :: Bool
is_fastf = case ClosureInfo -> Maybe (Int, ArgDescr)
closureFunInfo ClosureInfo
cl_info of
                 Just (Int
_, ArgGen Liveness
_) -> Bool
False
                 Maybe (Int, ArgDescr)
_otherwise         -> Bool
True

  Bool -> CmmExpr -> Int -> [LocalReg] -> FCode () -> FCode ()
entryHeapCheck' Bool
is_fastf CmmExpr
node Int
arity [LocalReg]
args FCode ()
code

-- | lower-level version for "GHC.Cmm.Parser"
entryHeapCheck' :: Bool           -- is a known function pattern
                -> CmmExpr        -- expression for the closure pointer
                -> Int            -- Arity -- not same as len args b/c of voids
                -> [LocalReg]     -- Non-void args (empty for thunk)
                -> FCode ()
                -> FCode ()
entryHeapCheck' :: Bool -> CmmExpr -> Int -> [LocalReg] -> FCode () -> FCode ()
entryHeapCheck' Bool
is_fastf CmmExpr
node Int
arity [LocalReg]
args FCode ()
code
  = do Profile
profile <- FCode Profile
getProfile
       let is_thunk :: Bool
is_thunk = Int
arity forall a. Eq a => a -> a -> Bool
== Int
0

           args' :: [CmmExpr]
args' = forall a b. (a -> b) -> [a] -> [b]
map (CmmReg -> CmmExpr
CmmReg forall b c a. (b -> c) -> (a -> b) -> a -> c
. LocalReg -> CmmReg
CmmLocal) [LocalReg]
args
           stg_gc_fun :: CmmExpr
stg_gc_fun    = CmmReg -> CmmExpr
CmmReg (GlobalReg -> CmmReg
CmmGlobal GlobalReg
GCFun)
           stg_gc_enter1 :: CmmExpr
stg_gc_enter1 = CmmReg -> CmmExpr
CmmReg (GlobalReg -> CmmReg
CmmGlobal GlobalReg
GCEnter1)

           {- Thunks:          jump stg_gc_enter_1

              Function (fast): call (NativeNode) stg_gc_fun(fun, args)

              Function (slow): call (slow) stg_gc_fun(fun, args)
           -}
           gc_call :: Int -> CmmAGraph
gc_call Int
upd
               | Bool
is_thunk
                 = Profile -> Convention -> CmmExpr -> [CmmExpr] -> Int -> CmmAGraph
mkJump Profile
profile Convention
NativeNodeCall CmmExpr
stg_gc_enter1 [CmmExpr
node] Int
upd

               | Bool
is_fastf
                 = Profile -> Convention -> CmmExpr -> [CmmExpr] -> Int -> CmmAGraph
mkJump Profile
profile Convention
NativeNodeCall CmmExpr
stg_gc_fun (CmmExpr
node forall a. a -> [a] -> [a]
: [CmmExpr]
args') Int
upd

               | Bool
otherwise
                 = Profile -> Convention -> CmmExpr -> [CmmExpr] -> Int -> CmmAGraph
mkJump Profile
profile Convention
Slow CmmExpr
stg_gc_fun (CmmExpr
node forall a. a -> [a] -> [a]
: [CmmExpr]
args') Int
upd

       Int
updfr_sz <- FCode Int
getUpdFrameOff

       BlockId
loop_id <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId
       BlockId -> FCode ()
emitLabel BlockId
loop_id
       forall a. Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck Bool
True Bool
True (Int -> CmmAGraph
gc_call Int
updfr_sz CmmAGraph -> CmmAGraph -> CmmAGraph
<*> BlockId -> CmmAGraph
mkBranch BlockId
loop_id) FCode ()
code

-- ------------------------------------------------------------
-- A heap/stack check in a case alternative


-- If there are multiple alts and we need to GC, but don't have a
-- continuation already (the scrut was simple), then we should
-- pre-generate the continuation.  (if there are multiple alts it is
-- always a canned GC point).

-- altHeapCheck:
-- If we have a return continuation,
--   then if it is a canned GC pattern,
--           then we do mkJumpReturnsTo
--           else we do a normal call to stg_gc_noregs
--   else if it is a canned GC pattern,
--           then generate the continuation and do mkCallReturnsTo
--           else we do a normal call to stg_gc_noregs

altHeapCheck :: [LocalReg] -> FCode a -> FCode a
altHeapCheck :: forall a. [LocalReg] -> FCode a -> FCode a
altHeapCheck [LocalReg]
regs FCode a
code = forall a. Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck Bool
False [LocalReg]
regs FCode a
code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck :: forall a. Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck Bool
checkYield [LocalReg]
regs FCode a
code = do
    Profile
profile <- FCode Profile
getProfile
    Platform
platform <- FCode Platform
getPlatform
    case Platform -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint Platform
platform [LocalReg]
regs of
      Maybe CmmExpr
Nothing -> forall a. Bool -> FCode a -> FCode a
genericGC Bool
checkYield FCode a
code
      Just CmmExpr
gc -> do
        BlockId
lret <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId
        let (Int
off, [GlobalReg]
_, CmmAGraph
copyin) = Profile
-> Convention
-> Area
-> [LocalReg]
-> [LocalReg]
-> (Int, [GlobalReg], CmmAGraph)
copyInOflow Profile
profile Convention
NativeReturn (BlockId -> Area
Young BlockId
lret) [LocalReg]
regs []
        BlockId
lcont <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId
        CmmTickScope
tscope <- FCode CmmTickScope
getTickScope
        BlockId -> CmmAGraphScoped -> FCode ()
emitOutOfLine BlockId
lret (CmmAGraph
copyin CmmAGraph -> CmmAGraph -> CmmAGraph
<*> BlockId -> CmmAGraph
mkBranch BlockId
lcont, CmmTickScope
tscope)
        BlockId -> FCode ()
emitLabel BlockId
lcont
        forall a.
Bool
-> Bool
-> CmmExpr
-> [LocalReg]
-> BlockId
-> Int
-> FCode a
-> FCode a
cannedGCReturnsTo Bool
checkYield Bool
False CmmExpr
gc [LocalReg]
regs BlockId
lret Int
off FCode a
code

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo :: forall a. [LocalReg] -> BlockId -> Int -> FCode a -> FCode a
altHeapCheckReturnsTo [LocalReg]
regs BlockId
lret Int
off FCode a
code
  = do Platform
platform <- FCode Platform
getPlatform
       case Platform -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint Platform
platform [LocalReg]
regs of
           Maybe CmmExpr
Nothing -> forall a. Bool -> FCode a -> FCode a
genericGC Bool
False FCode a
code
           Just CmmExpr
gc -> forall a.
Bool
-> Bool
-> CmmExpr
-> [LocalReg]
-> BlockId
-> Int
-> FCode a
-> FCode a
cannedGCReturnsTo Bool
False Bool
True CmmExpr
gc [LocalReg]
regs BlockId
lret Int
off FCode a
code

-- noEscapeHeapCheck is implemented identically to altHeapCheck (which
-- is more efficient), but cannot be optimized away in the non-allocating
-- case because it may occur in a loop
noEscapeHeapCheck :: [LocalReg] -> FCode a -> FCode a
noEscapeHeapCheck :: forall a. [LocalReg] -> FCode a -> FCode a
noEscapeHeapCheck = forall a. Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck Bool
True

cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
                  -> FCode a
                  -> FCode a
cannedGCReturnsTo :: forall a.
Bool
-> Bool
-> CmmExpr
-> [LocalReg]
-> BlockId
-> Int
-> FCode a
-> FCode a
cannedGCReturnsTo Bool
checkYield Bool
cont_on_stack CmmExpr
gc [LocalReg]
regs BlockId
lret Int
off FCode a
code
  = do Profile
profile <- FCode Profile
getProfile
       Int
updfr_sz <- FCode Int
getUpdFrameOff
       forall a. Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck Bool
False Bool
checkYield (Profile -> CmmExpr -> Int -> CmmAGraph
gc_call Profile
profile CmmExpr
gc Int
updfr_sz) FCode a
code
  where
    reg_exprs :: [CmmExpr]
reg_exprs = forall a b. (a -> b) -> [a] -> [b]
map (CmmReg -> CmmExpr
CmmReg forall b c a. (b -> c) -> (a -> b) -> a -> c
. LocalReg -> CmmReg
CmmLocal) [LocalReg]
regs
      -- Note [stg_gc arguments]

      -- NB. we use the NativeReturn convention for passing arguments
      -- to the canned heap-check routines, because we are in a case
      -- alternative and hence the [LocalReg] was passed to us in the
      -- NativeReturn convention.
    gc_call :: Profile -> CmmExpr -> Int -> CmmAGraph
gc_call Profile
profile CmmExpr
label Int
sp
      | Bool
cont_on_stack
      = Profile
-> CmmExpr
-> Convention
-> [CmmExpr]
-> BlockId
-> Int
-> Int
-> CmmAGraph
mkJumpReturnsTo Profile
profile CmmExpr
label Convention
NativeReturn [CmmExpr]
reg_exprs BlockId
lret Int
off Int
sp
      | Bool
otherwise
      = Profile
-> CmmExpr
-> Convention
-> [CmmExpr]
-> BlockId
-> Int
-> Int
-> [CmmExpr]
-> CmmAGraph
mkCallReturnsTo Profile
profile CmmExpr
label Convention
NativeReturn [CmmExpr]
reg_exprs BlockId
lret Int
off Int
sp []

genericGC :: Bool -> FCode a -> FCode a
genericGC :: forall a. Bool -> FCode a -> FCode a
genericGC Bool
checkYield FCode a
code
  = do Int
updfr_sz <- FCode Int
getUpdFrameOff
       BlockId
lretry <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId
       BlockId -> FCode ()
emitLabel BlockId
lretry
       CmmAGraph
call <- CmmExpr
-> (Convention, Convention)
-> [LocalReg]
-> [CmmExpr]
-> Int
-> [CmmExpr]
-> FCode CmmAGraph
mkCall CmmExpr
generic_gc (Convention
GC, Convention
GC) [] [] Int
updfr_sz []
       forall a. Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck Bool
False Bool
checkYield (CmmAGraph
call CmmAGraph -> CmmAGraph -> CmmAGraph
<*> BlockId -> CmmAGraph
mkBranch BlockId
lretry) FCode a
code

cannedGCEntryPoint :: Platform -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint :: Platform -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint Platform
platform [LocalReg]
regs
  = case forall a b. (a -> b) -> [a] -> [b]
map LocalReg -> CmmType
localRegType [LocalReg]
regs of
      []  -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_noregs")
      [CmmType
ty]
          | CmmType -> Bool
isGcPtrType CmmType
ty -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_unpt_r1")
          | CmmType -> Bool
isFloatType CmmType
ty -> case Width
width of
                                  Width
W32       -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_f1")
                                  Width
W64       -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_d1")
                                  Width
_         -> forall a. Maybe a
Nothing

          | Width
width forall a. Eq a => a -> a -> Bool
== Platform -> Width
wordWidth Platform
platform -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_unbx_r1")
          | Width
width forall a. Eq a => a -> a -> Bool
== Width
W64                -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_l1")
          | Bool
otherwise                   -> forall a. Maybe a
Nothing
          where
              width :: Width
width = CmmType -> Width
typeWidth CmmType
ty
      [CmmType
ty1,CmmType
ty2]
          |  CmmType -> Bool
isGcPtrType CmmType
ty1
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty2 -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_pp")
      [CmmType
ty1,CmmType
ty2,CmmType
ty3]
          |  CmmType -> Bool
isGcPtrType CmmType
ty1
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty2
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty3 -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_ppp")
      [CmmType
ty1,CmmType
ty2,CmmType
ty3,CmmType
ty4]
          |  CmmType -> Bool
isGcPtrType CmmType
ty1
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty2
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty3
          Bool -> Bool -> Bool
&& CmmType -> Bool
isGcPtrType CmmType
ty4 -> forall a. a -> Maybe a
Just (String -> CmmExpr
mkGcLabel String
"stg_gc_pppp")
      [CmmType]
_otherwise -> forall a. Maybe a
Nothing

-- Note [stg_gc arguments]
-- ~~~~~~~~~~~~~~~~~~~~~~~
-- It might seem that we could avoid passing the arguments to the
-- stg_gc function, because they are already in the right registers.
-- While this is usually the case, it isn't always.  Sometimes the
-- code generator has cleverly avoided the eval in a case, e.g. in
-- ffi/should_run/4221.hs we found
--
--   case a_r1mb of z
--     FunPtr x y -> ...
--
-- where a_r1mb is bound a top-level constructor, and is known to be
-- evaluated.  The codegen just assigns x, y and z, and continues;
-- R1 is never assigned.
--
-- So we'll have to rely on optimisations to eliminatethese
-- assignments where possible.


-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
generic_gc :: CmmExpr
generic_gc = String -> CmmExpr
mkGcLabel String
"stg_gc_noregs"

-- | Create a CLabel for calling a garbage collector entry point
mkGcLabel :: String -> CmmExpr
mkGcLabel :: String -> CmmExpr
mkGcLabel String
s = CmmLit -> CmmExpr
CmmLit (CLabel -> CmmLit
CmmLabel (UnitId -> FastString -> CLabel
mkCmmCodeLabel UnitId
rtsUnitId (String -> FastString
fsLit String
s)))

-------------------------------
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck :: forall a. Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck Bool
checkStack Bool
checkYield CmmAGraph
do_gc FCode a
code
  = forall a. (Int -> FCode a) -> FCode a
getHeapUsage forall a b. (a -> b) -> a -> b
$ \ Int
hpHw ->
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
    do  { Platform
platform <- FCode Platform
getPlatform
        ; let mb_alloc_bytes :: Maybe CmmExpr
mb_alloc_bytes
                 | Int
hpHw forall a. Ord a => a -> a -> Bool
> Int
mBLOCK_SIZE = forall a. String -> a
sorry forall a b. (a -> b) -> a -> b
$ [String] -> String
unlines
                    [String
" Trying to allocate more than "forall a. [a] -> [a] -> [a]
++forall a. Show a => a -> String
show Int
mBLOCK_SIZEforall a. [a] -> [a] -> [a]
++String
" bytes.",
                     String
"",
                     String
"This is currently not possible due to a limitation of GHC's code generator.",
                     String
"See https://gitlab.haskell.org/ghc/ghc/issues/4505 for details.",
                     String
"Suggestion: read data from a file instead of having large static data",
                     String
"structures in code."]
                 | Int
hpHw forall a. Ord a => a -> a -> Bool
> Int
0  = forall a. a -> Maybe a
Just (Platform -> Int -> CmmExpr
mkIntExpr Platform
platform (Int
hpHw forall a. Num a => a -> a -> a
* (Platform -> Int
platformWordSizeInBytes Platform
platform)))
                 | Bool
otherwise = forall a. Maybe a
Nothing
                 where
                  constants :: PlatformConstants
constants = Platform -> PlatformConstants
platformConstants Platform
platform
                  bLOCK_SIZE_W :: Int
bLOCK_SIZE_W = PlatformConstants -> Int
pc_BLOCK_SIZE (Platform -> PlatformConstants
platformConstants Platform
platform) forall a. Integral a => a -> a -> a
`quot` Platform -> Int
platformWordSizeInBytes Platform
platform
                  mBLOCK_SIZE :: Int
mBLOCK_SIZE = PlatformConstants -> Int
pc_BLOCKS_PER_MBLOCK PlatformConstants
constants forall a. Num a => a -> a -> a
* Int
bLOCK_SIZE_W
              stk_hwm :: Maybe CmmExpr
stk_hwm | Bool
checkStack = forall a. a -> Maybe a
Just (CmmLit -> CmmExpr
CmmLit CmmLit
CmmHighStackMark)
                      | Bool
otherwise  = forall a. Maybe a
Nothing
        ; FCode () -> FCode ()
codeOnly forall a b. (a -> b) -> a -> b
$ Maybe CmmExpr -> Bool -> Maybe CmmExpr -> CmmAGraph -> FCode ()
do_checks Maybe CmmExpr
stk_hwm Bool
checkYield Maybe CmmExpr
mb_alloc_bytes CmmAGraph
do_gc
        ; Bool -> Int -> FCode ()
tickyAllocHeap Bool
True Int
hpHw
        ; Int -> FCode ()
setRealHp Int
hpHw
        ; FCode a
code }

heapStackCheckGen :: Maybe CmmExpr -> Maybe CmmExpr -> FCode ()
heapStackCheckGen :: Maybe CmmExpr -> Maybe CmmExpr -> FCode ()
heapStackCheckGen Maybe CmmExpr
stk_hwm Maybe CmmExpr
mb_bytes
  = do Int
updfr_sz <- FCode Int
getUpdFrameOff
       BlockId
lretry <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId
       BlockId -> FCode ()
emitLabel BlockId
lretry
       CmmAGraph
call <- CmmExpr
-> (Convention, Convention)
-> [LocalReg]
-> [CmmExpr]
-> Int
-> [CmmExpr]
-> FCode CmmAGraph
mkCall CmmExpr
generic_gc (Convention
GC, Convention
GC) [] [] Int
updfr_sz []
       Maybe CmmExpr -> Bool -> Maybe CmmExpr -> CmmAGraph -> FCode ()
do_checks Maybe CmmExpr
stk_hwm Bool
False Maybe CmmExpr
mb_bytes (CmmAGraph
call CmmAGraph -> CmmAGraph -> CmmAGraph
<*> BlockId -> CmmAGraph
mkBranch BlockId
lretry)

-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
-- When compiling a function we can determine how much stack space it
-- will use. We therefore need to perform only a single stack check at
-- the beginning of a function to see if we have enough stack space.
--
-- The check boils down to comparing Sp-N with SpLim, where N is the
-- amount of stack space needed (see Note [Stack usage] below).  *BUT*
-- at this stage of the pipeline we are not supposed to refer to Sp
-- itself, because the stack is not yet manifest, so we don't quite
-- know where Sp pointing.

-- So instead of referring directly to Sp - as we used to do in the
-- past - the code generator uses (old + 0) in the stack check. That
-- is the address of the first word of the old area, so if we add N
-- we'll get the address of highest used word.
--
-- This makes the check robust.  For example, while we need to perform
-- only one stack check for each function, we could in theory place
-- more stack checks later in the function. They would be redundant,
-- but not incorrect (in a sense that they should not change program
-- behaviour). We need to make sure however that a stack check
-- inserted after incrementing the stack pointer checks for a
-- respectively smaller stack space. This would not be the case if the
-- code generator produced direct references to Sp. By referencing
-- (old + 0) we make sure that we always check for a correct amount of
-- stack: when converting (old + 0) to Sp the stack layout phase takes
-- into account changes already made to stack pointer. The idea for
-- this change came from observations made while debugging #8275.

-- Note [Stack usage]
-- ~~~~~~~~~~~~~~~~~~
-- At the moment we convert from STG to Cmm we don't know N, the
-- number of bytes of stack that the function will use, so we use a
-- special late-bound CmmLit, namely
--       CmmHighStackMark
-- to stand for the number of bytes needed. When the stack is made
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
-- The (Maybe CmmExpr) passed to do_checks is usually
--     Just (CmmLit CmmHighStackMark)
-- but can also (in certain hand-written RTS functions)
--     Just (CmmLit 8)  or some other fixed valuet
-- If it is Nothing, we don't generate a stack check at all.

do_checks :: Maybe CmmExpr    -- Should we check the stack?
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
          -> Maybe CmmExpr    -- Heap headroom (bytes)
          -> CmmAGraph        -- What to do on failure
          -> FCode ()
do_checks :: Maybe CmmExpr -> Bool -> Maybe CmmExpr -> CmmAGraph -> FCode ()
do_checks Maybe CmmExpr
mb_stk_hwm Bool
checkYield Maybe CmmExpr
mb_alloc_lit CmmAGraph
do_gc = do
  Bool
omit_yields <- StgToCmmConfig -> Bool
stgToCmmOmitYields forall (f :: * -> *) a b. Functor f => (a -> b) -> f a -> f b
<$> FCode StgToCmmConfig
getStgToCmmConfig
  Platform
platform    <- FCode Platform
getPlatform
  BlockId
gc_id       <- forall (m :: * -> *). MonadUnique m => m BlockId
newBlockId

  let
    Just CmmExpr
alloc_lit = Maybe CmmExpr
mb_alloc_lit

    bump_hp :: CmmExpr
bump_hp   = Platform -> CmmExpr -> CmmExpr -> CmmExpr
cmmOffsetExprB Platform
platform CmmExpr
hpExpr CmmExpr
alloc_lit

    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
    sp_oflo :: CmmExpr -> CmmExpr
sp_oflo CmmExpr
sp_hwm =
         MachOp -> [CmmExpr] -> CmmExpr
CmmMachOp (Platform -> MachOp
mo_wordULt Platform
platform)
                  [MachOp -> [CmmExpr] -> CmmExpr
CmmMachOp (Width -> MachOp
MO_Sub (CmmType -> Width
typeWidth (Platform -> CmmReg -> CmmType
cmmRegType Platform
platform CmmReg
spReg)))
                             [Area -> Int -> CmmExpr
CmmStackSlot Area
Old Int
0, CmmExpr
sp_hwm],
                   CmmReg -> CmmExpr
CmmReg CmmReg
spLimReg]

    -- Hp overflow if (Hp > HpLim)
    -- (Hp has been incremented by now)
    -- HpLim points to the LAST WORD of valid allocation space.
    hp_oflo :: CmmExpr
hp_oflo = MachOp -> [CmmExpr] -> CmmExpr
CmmMachOp (Platform -> MachOp
mo_wordUGt Platform
platform) [CmmExpr
hpExpr, CmmExpr
hpLimExpr]

    alloc_n :: CmmAGraph
alloc_n = CmmReg -> CmmExpr -> CmmAGraph
mkAssign CmmReg
hpAllocReg CmmExpr
alloc_lit

  case Maybe CmmExpr
mb_stk_hwm of
    Maybe CmmExpr
Nothing -> forall (m :: * -> *) a. Monad m => a -> m a
return ()
    Just CmmExpr
stk_hwm -> FCode ()
tickyStackCheck
      forall (m :: * -> *) a b. Monad m => m a -> m b -> m b
>> (CmmAGraph -> FCode ()
emit forall (m :: * -> *) a b. Monad m => (a -> m b) -> m a -> m b
=<< CmmExpr -> BlockId -> Maybe Bool -> FCode CmmAGraph
mkCmmIfGoto' (CmmExpr -> CmmExpr
sp_oflo CmmExpr
stk_hwm) BlockId
gc_id (forall a. a -> Maybe a
Just Bool
False) )

  -- Emit new label that might potentially be a header
  -- of a self-recursive tail call.
  -- See Note [Self-recursive loop header].
  Maybe SelfLoopInfo
self_loop_info <- FCode (Maybe SelfLoopInfo)
getSelfLoop
  case Maybe SelfLoopInfo
self_loop_info of
    Just (Id
_, BlockId
loop_header_id, [LocalReg]
_)
        | Bool
checkYield Bool -> Bool -> Bool
&& forall a. Maybe a -> Bool
isJust Maybe CmmExpr
mb_stk_hwm -> BlockId -> FCode ()
emitLabel BlockId
loop_header_id
    Maybe SelfLoopInfo
_otherwise -> forall (m :: * -> *) a. Monad m => a -> m a
return ()

  if forall a. Maybe a -> Bool
isJust Maybe CmmExpr
mb_alloc_lit
    then do
     FCode ()
tickyHeapCheck
     CmmReg -> CmmExpr -> FCode ()
emitAssign CmmReg
hpReg CmmExpr
bump_hp
     CmmAGraph -> FCode ()
emit forall (m :: * -> *) a b. Monad m => (a -> m b) -> m a -> m b
=<< CmmExpr -> CmmAGraph -> Maybe Bool -> FCode CmmAGraph
mkCmmIfThen' CmmExpr
hp_oflo (CmmAGraph
alloc_n CmmAGraph -> CmmAGraph -> CmmAGraph
<*> BlockId -> CmmAGraph
mkBranch BlockId
gc_id) (forall a. a -> Maybe a
Just Bool
False)
    else
      forall (f :: * -> *). Applicative f => Bool -> f () -> f ()
when (Bool
checkYield Bool -> Bool -> Bool
&& Bool -> Bool
not Bool
omit_yields) forall a b. (a -> b) -> a -> b
$ do
         -- Yielding if HpLim == 0
         let yielding :: CmmExpr
yielding = MachOp -> [CmmExpr] -> CmmExpr
CmmMachOp (Platform -> MachOp
mo_wordEq Platform
platform)
                                  [CmmReg -> CmmExpr
CmmReg CmmReg
hpLimReg,
                                   CmmLit -> CmmExpr
CmmLit (Platform -> CmmLit
zeroCLit Platform
platform)]
         CmmAGraph -> FCode ()
emit forall (m :: * -> *) a b. Monad m => (a -> m b) -> m a -> m b
=<< CmmExpr -> BlockId -> Maybe Bool -> FCode CmmAGraph
mkCmmIfGoto' CmmExpr
yielding BlockId
gc_id (forall a. a -> Maybe a
Just Bool
False)

  CmmTickScope
tscope <- FCode CmmTickScope
getTickScope
  BlockId -> CmmAGraphScoped -> FCode ()
emitOutOfLine BlockId
gc_id
   (CmmAGraph
do_gc, CmmTickScope
tscope) -- this is expected to jump back somewhere

                -- Test for stack pointer exhaustion, then
                -- bump heap pointer, and test for heap exhaustion
                -- Note that we don't move the heap pointer unless the
                -- stack check succeeds.  Otherwise we might end up
                -- with slop at the end of the current block, which can
                -- confuse the LDV profiler.

-- Note [Self-recursive loop header]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- Self-recursive loop header is required by loopification optimization (See
-- Note [Self-recursive tail calls] in GHC.StgToCmm.Expr). We emit it if:
--
--  1. There is information about self-loop in the FCode environment. We don't
--     check the binder (first component of the self_loop_info) because we are
--     certain that if the self-loop info is present then we are compiling the
--     binder body. Reason: the only possible way to get here with the
--     self_loop_info present is from closureCodeBody.
--
--  2. checkYield && isJust mb_stk_hwm. checkYield tells us that it is possible
--     to preempt the heap check (see #367 for motivation behind this check). It
--     is True for heap checks placed at the entry to a function and
--     let-no-escape heap checks but false for other heap checks (eg. in case
--     alternatives or created from hand-written high-level Cmm). The second
--     check (isJust mb_stk_hwm) is true for heap checks at the entry to a
--     function and some heap checks created in hand-written Cmm. Otherwise it
--     is Nothing. In other words the only situation when both conditions are
--     true is when compiling stack and heap checks at the entry to a
--     function. This is the only situation when we want to emit a self-loop
--     label.