{-# LANGUAGE BangPatterns, RecordWildCards, GADTs #-} module CmmLayoutStack ( cmmLayoutStack, setInfoTableStackMap ) where import GhcPrelude hiding ((<*>)) import GHC.StgToCmm.Utils ( callerSaveVolatileRegs, newTemp ) -- XXX layering violation import GHC.StgToCmm.Foreign ( saveThreadState, loadThreadState ) -- XXX layering violation import BasicTypes import Cmm import CmmInfo import BlockId import CLabel import CmmUtils import MkGraph import ForeignCall import CmmLive import CmmProcPoint import SMRep import Hoopl.Block import Hoopl.Collections import Hoopl.Dataflow import Hoopl.Graph import Hoopl.Label import UniqSupply import Maybes import UniqFM import Util import DynFlags import FastString import Outputable hiding ( isEmpty ) import qualified Data.Set as Set import Control.Monad.Fix import Data.Array as Array import Data.Bits import Data.List (nub) import TyCon (PrimRep (..)) {- Note [Stack Layout] The job of this pass is to - replace references to abstract stack Areas with fixed offsets from Sp. - replace the CmmHighStackMark constant used in the stack check with the maximum stack usage of the proc. - save any variables that are live across a call, and reload them as necessary. Before stack allocation, local variables remain live across native calls (CmmCall{ cmm_cont = Just _ }), and after stack allocation local variables are clobbered by native calls. We want to do stack allocation so that as far as possible - stack use is minimized, and - unnecessary stack saves and loads are avoided. The algorithm we use is a variant of linear-scan register allocation, where the stack is our register file. We proceed in two passes, see Note [Two pass approach] for why they are not easy to merge into one. Pass 1: - First, we do a liveness analysis, which annotates every block with the variables live on entry to the block. - We traverse blocks in reverse postorder DFS; that is, we visit at least one predecessor of a block before the block itself. The stack layout flowing from the predecessor of the block will determine the stack layout on entry to the block. - We maintain a data structure Map Label StackMap which describes the contents of the stack and the stack pointer on entry to each block that is a successor of a block that we have visited. - For each block we visit: - Look up the StackMap for this block. - If this block is a proc point (or a call continuation, if we aren't splitting proc points), we need to reload all the live variables from the stack - but this is done in Pass 2, which calculates more precise liveness information (see description of Pass 2). - Walk forwards through the instructions: - At an assignment x = Sp[loc] - Record the fact that Sp[loc] contains x, so that we won't need to save x if it ever needs to be spilled. - At an assignment x = E - If x was previously on the stack, it isn't any more - At the last node, if it is a call or a jump to a proc point - Lay out the stack frame for the call (see setupStackFrame) - emit instructions to save all the live variables - Remember the StackMaps for all the successors - emit an instruction to adjust Sp - If the last node is a branch, then the current StackMap is the StackMap for the successors. - Manifest Sp: replace references to stack areas in this block with real Sp offsets. We cannot do this until we have laid out the stack area for the successors above. In this phase we also eliminate redundant stores to the stack; see elimStackStores. - There is one important gotcha: sometimes we'll encounter a control transfer to a block that we've already processed (a join point), and in that case we might need to rearrange the stack to match what the block is expecting. (exactly the same as in linear-scan register allocation, except here we have the luxury of an infinite supply of temporary variables). - Finally, we update the magic CmmHighStackMark constant with the stack usage of the function, and eliminate the whole stack check if there was no stack use. (in fact this is done as part of the main traversal, by feeding the high-water-mark output back in as an input. I hate cyclic programming, but it's just too convenient sometimes.) There are plenty of tricky details: update frames, proc points, return addresses, foreign calls, and some ad-hoc optimisations that are convenient to do here and effective in common cases. Comments in the code below explain these. Pass 2: - Calculate live registers, but taking into account that nothing is live at the entry to a proc point. - At each proc point and call continuation insert reloads of live registers from the stack (they were saved by Pass 1). Note [Two pass approach] The main reason for Pass 2 is being able to insert only the reloads that are needed and the fact that the two passes need different liveness information. Let's consider an example: ..... \ / D <- proc point / \ E F \ / G <- proc point | X Pass 1 needs liveness assuming that local variables are preserved across calls. This is important because it needs to save any local registers to the stack (e.g., if register a is used in block X, it must be saved before any native call). However, for Pass 2, where we want to reload registers from stack (in a proc point), this is overly conservative and would lead us to generate reloads in D for things used in X, even though we're going to generate reloads in G anyway (since it's also a proc point). So Pass 2 calculates liveness knowing that nothing is live at the entry to a proc point. This means that in D we only need to reload things used in E or F. This can be quite important, for an extreme example see testcase for #3294. Merging the two passes is not trivial - Pass 2 is a backward rewrite and Pass 1 is a forward one. Furthermore, Pass 1 is creating code that uses local registers (saving them before a call), which the liveness analysis for Pass 2 must see to be correct. -} -- All stack locations are expressed as positive byte offsets from the -- "base", which is defined to be the address above the return address -- on the stack on entry to this CmmProc. -- -- Lower addresses have higher StackLocs. -- type StackLoc = ByteOff {- A StackMap describes the stack at any given point. At a continuation it has a particular layout, like this: | | <- base |-------------| | ret0 | <- base + 8 |-------------| . upd frame . <- base + sm_ret_off |-------------| | | . vars . . (live/dead) . | | <- base + sm_sp - sm_args |-------------| | ret1 | . ret vals . <- base + sm_sp (<--- Sp points here) |-------------| Why do we include the final return address (ret0) in our stack map? I have absolutely no idea, but it seems to be done that way consistently in the rest of the code generator, so I played along here. --SDM Note that we will be constructing an info table for the continuation (ret1), which needs to describe the stack down to, but not including, the update frame (or ret0, if there is no update frame). -} data StackMap = StackMap { sm_sp :: StackLoc -- ^ the offset of Sp relative to the base on entry -- to this block. , sm_args :: ByteOff -- ^ the number of bytes of arguments in the area for this block -- Defn: the offset of young(L) relative to the base is given by -- (sm_sp - sm_args) of the StackMap for block L. , sm_ret_off :: ByteOff -- ^ Number of words of stack that we do not describe with an info -- table, because it contains an update frame. , sm_regs :: UniqFM (LocalReg,StackLoc) -- ^ regs on the stack } instance Outputable StackMap where ppr StackMap{..} = text "Sp = " <> int sm_sp $$ text "sm_args = " <> int sm_args $$ text "sm_ret_off = " <> int sm_ret_off $$ text "sm_regs = " <> pprUFM sm_regs ppr cmmLayoutStack :: DynFlags -> ProcPointSet -> ByteOff -> CmmGraph -> UniqSM (CmmGraph, LabelMap StackMap) cmmLayoutStack dflags procpoints entry_args graph@(CmmGraph { g_entry = entry }) = do -- We need liveness info. Dead assignments are removed later -- by the sinking pass. let liveness = cmmLocalLiveness dflags graph blocks = revPostorder graph (final_stackmaps, _final_high_sp, new_blocks) <- mfix $ \ ~(rec_stackmaps, rec_high_sp, _new_blocks) -> layout dflags procpoints liveness entry entry_args rec_stackmaps rec_high_sp blocks blocks_with_reloads <- insertReloadsAsNeeded dflags procpoints final_stackmaps entry new_blocks new_blocks' <- mapM (lowerSafeForeignCall dflags) blocks_with_reloads return (ofBlockList entry new_blocks', final_stackmaps) -- ----------------------------------------------------------------------------- -- Pass 1 -- ----------------------------------------------------------------------------- layout :: DynFlags -> LabelSet -- proc points -> LabelMap CmmLocalLive -- liveness -> BlockId -- entry -> ByteOff -- stack args on entry -> LabelMap StackMap -- [final] stack maps -> ByteOff -- [final] Sp high water mark -> [CmmBlock] -- [in] blocks -> UniqSM ( LabelMap StackMap -- [out] stack maps , ByteOff -- [out] Sp high water mark , [CmmBlock] -- [out] new blocks ) layout dflags procpoints liveness entry entry_args final_stackmaps final_sp_high blocks = go blocks init_stackmap entry_args [] where (updfr, cont_info) = collectContInfo blocks init_stackmap = mapSingleton entry StackMap{ sm_sp = entry_args , sm_args = entry_args , sm_ret_off = updfr , sm_regs = emptyUFM } go [] acc_stackmaps acc_hwm acc_blocks = return (acc_stackmaps, acc_hwm, acc_blocks) go (b0 : bs) acc_stackmaps acc_hwm acc_blocks = do let (entry0@(CmmEntry entry_lbl tscope), middle0, last0) = blockSplit b0 let stack0@StackMap { sm_sp = sp0 } = mapFindWithDefault (pprPanic "no stack map for" (ppr entry_lbl)) entry_lbl acc_stackmaps -- (a) Update the stack map to include the effects of -- assignments in this block let stack1 = foldBlockNodesF (procMiddle acc_stackmaps) middle0 stack0 -- (b) Look at the last node and if we are making a call or -- jumping to a proc point, we must save the live -- variables, adjust Sp, and construct the StackMaps for -- each of the successor blocks. See handleLastNode for -- details. (middle1, sp_off, last1, fixup_blocks, out) <- handleLastNode dflags procpoints liveness cont_info acc_stackmaps stack1 tscope middle0 last0 -- (c) Manifest Sp: run over the nodes in the block and replace -- CmmStackSlot with CmmLoad from Sp with a concrete offset. -- -- our block: -- middle0 -- the original middle nodes -- middle1 -- live variable saves from handleLastNode -- Sp = Sp + sp_off -- Sp adjustment goes here -- last1 -- the last node -- let middle_pre = blockToList $ foldl' blockSnoc middle0 middle1 let final_blocks = manifestSp dflags final_stackmaps stack0 sp0 final_sp_high entry0 middle_pre sp_off last1 fixup_blocks let acc_stackmaps' = mapUnion acc_stackmaps out -- If this block jumps to the GC, then we do not take its -- stack usage into account for the high-water mark. -- Otherwise, if the only stack usage is in the stack-check -- failure block itself, we will do a redundant stack -- check. The stack has a buffer designed to accommodate -- the largest amount of stack needed for calling the GC. -- this_sp_hwm | isGcJump last0 = 0 | otherwise = sp0 - sp_off hwm' = maximum (acc_hwm : this_sp_hwm : map sm_sp (mapElems out)) go bs acc_stackmaps' hwm' (final_blocks ++ acc_blocks) -- ----------------------------------------------------------------------------- -- Not foolproof, but GCFun is the culprit we most want to catch isGcJump :: CmmNode O C -> Bool isGcJump (CmmCall { cml_target = CmmReg (CmmGlobal l) }) = l == GCFun || l == GCEnter1 isGcJump _something_else = False -- ----------------------------------------------------------------------------- -- This doesn't seem right somehow. We need to find out whether this -- proc will push some update frame material at some point, so that we -- can avoid using that area of the stack for spilling. The -- updfr_space field of the CmmProc *should* tell us, but it doesn't -- (I think maybe it gets filled in later when we do proc-point -- splitting). -- -- So we'll just take the max of all the cml_ret_offs. This could be -- unnecessarily pessimistic, but probably not in the code we -- generate. collectContInfo :: [CmmBlock] -> (ByteOff, LabelMap ByteOff) collectContInfo blocks = (maximum ret_offs, mapFromList (catMaybes mb_argss)) where (mb_argss, ret_offs) = mapAndUnzip get_cont blocks get_cont :: Block CmmNode x C -> (Maybe (Label, ByteOff), ByteOff) get_cont b = case lastNode b of CmmCall { cml_cont = Just l, .. } -> (Just (l, cml_ret_args), cml_ret_off) CmmForeignCall { .. } -> (Just (succ, ret_args), ret_off) _other -> (Nothing, 0) -- ----------------------------------------------------------------------------- -- Updating the StackMap from middle nodes -- Look for loads from stack slots, and update the StackMap. This is -- purely for optimisation reasons, so that we can avoid saving a -- variable back to a different stack slot if it is already on the -- stack. -- -- This happens a lot: for example when function arguments are passed -- on the stack and need to be immediately saved across a call, we -- want to just leave them where they are on the stack. -- procMiddle :: LabelMap StackMap -> CmmNode e x -> StackMap -> StackMap procMiddle stackmaps node sm = case node of CmmAssign (CmmLocal r) (CmmLoad (CmmStackSlot area off) _) -> sm { sm_regs = addToUFM (sm_regs sm) r (r,loc) } where loc = getStackLoc area off stackmaps CmmAssign (CmmLocal r) _other -> sm { sm_regs = delFromUFM (sm_regs sm) r } _other -> sm getStackLoc :: Area -> ByteOff -> LabelMap StackMap -> StackLoc getStackLoc Old n _ = n getStackLoc (Young l) n stackmaps = case mapLookup l stackmaps of Nothing -> pprPanic "getStackLoc" (ppr l) Just sm -> sm_sp sm - sm_args sm + n -- ----------------------------------------------------------------------------- -- Handling stack allocation for a last node -- We take a single last node and turn it into: -- -- C1 (some statements) -- Sp = Sp + N -- C2 (some more statements) -- call f() -- the actual last node -- -- plus possibly some more blocks (we may have to add some fixup code -- between the last node and the continuation). -- -- C1: is the code for saving the variables across this last node onto -- the stack, if the continuation is a call or jumps to a proc point. -- -- C2: if the last node is a safe foreign call, we have to inject some -- extra code that goes *after* the Sp adjustment. handleLastNode :: DynFlags -> ProcPointSet -> LabelMap CmmLocalLive -> LabelMap ByteOff -> LabelMap StackMap -> StackMap -> CmmTickScope -> Block CmmNode O O -> CmmNode O C -> UniqSM ( [CmmNode O O] -- nodes to go *before* the Sp adjustment , ByteOff -- amount to adjust Sp , CmmNode O C -- new last node , [CmmBlock] -- new blocks , LabelMap StackMap -- stackmaps for the continuations ) handleLastNode dflags procpoints liveness cont_info stackmaps stack0@StackMap { sm_sp = sp0 } tscp middle last = case last of -- At each return / tail call, -- adjust Sp to point to the last argument pushed, which -- is cml_args, after popping any other junk from the stack. CmmCall{ cml_cont = Nothing, .. } -> do let sp_off = sp0 - cml_args return ([], sp_off, last, [], mapEmpty) -- At each CmmCall with a continuation: CmmCall{ cml_cont = Just cont_lbl, .. } -> return $ lastCall cont_lbl cml_args cml_ret_args cml_ret_off CmmForeignCall{ succ = cont_lbl, .. } -> do return $ lastCall cont_lbl (wORD_SIZE dflags) ret_args ret_off -- one word of args: the return address CmmBranch {} -> handleBranches CmmCondBranch {} -> handleBranches CmmSwitch {} -> handleBranches where -- Calls and ForeignCalls are handled the same way: lastCall :: BlockId -> ByteOff -> ByteOff -> ByteOff -> ( [CmmNode O O] , ByteOff , CmmNode O C , [CmmBlock] , LabelMap StackMap ) lastCall lbl cml_args cml_ret_args cml_ret_off = ( assignments , spOffsetForCall sp0 cont_stack cml_args , last , [] -- no new blocks , mapSingleton lbl cont_stack ) where (assignments, cont_stack) = prepareStack lbl cml_ret_args cml_ret_off prepareStack lbl cml_ret_args cml_ret_off | Just cont_stack <- mapLookup lbl stackmaps -- If we have already seen this continuation before, then -- we just have to make the stack look the same: = (fixupStack stack0 cont_stack, cont_stack) -- Otherwise, we have to allocate the stack frame | otherwise = (save_assignments, new_cont_stack) where (new_cont_stack, save_assignments) = setupStackFrame dflags lbl liveness cml_ret_off cml_ret_args stack0 -- For other last nodes (branches), if any of the targets is a -- proc point, we have to set up the stack to match what the proc -- point is expecting. -- handleBranches :: UniqSM ( [CmmNode O O] , ByteOff , CmmNode O C , [CmmBlock] , LabelMap StackMap ) handleBranches -- Note [diamond proc point] | Just l <- futureContinuation middle , (nub $ filter (`setMember` procpoints) $ successors last) == [l] = do let cont_args = mapFindWithDefault 0 l cont_info (assigs, cont_stack) = prepareStack l cont_args (sm_ret_off stack0) out = mapFromList [ (l', cont_stack) | l' <- successors last ] return ( assigs , spOffsetForCall sp0 cont_stack (wORD_SIZE dflags) , last , [] , out) | otherwise = do pps <- mapM handleBranch (successors last) let lbl_map :: LabelMap Label lbl_map = mapFromList [ (l,tmp) | (l,tmp,_,_) <- pps ] fix_lbl l = mapFindWithDefault l l lbl_map return ( [] , 0 , mapSuccessors fix_lbl last , concat [ blk | (_,_,_,blk) <- pps ] , mapFromList [ (l, sm) | (l,_,sm,_) <- pps ] ) -- For each successor of this block handleBranch :: BlockId -> UniqSM (BlockId, BlockId, StackMap, [CmmBlock]) handleBranch l -- (a) if the successor already has a stackmap, we need to -- shuffle the current stack to make it look the same. -- We have to insert a new block to make this happen. | Just stack2 <- mapLookup l stackmaps = do let assigs = fixupStack stack0 stack2 (tmp_lbl, block) <- makeFixupBlock dflags sp0 l stack2 tscp assigs return (l, tmp_lbl, stack2, block) -- (b) if the successor is a proc point, save everything -- on the stack. | l `setMember` procpoints = do let cont_args = mapFindWithDefault 0 l cont_info (stack2, assigs) = setupStackFrame dflags l liveness (sm_ret_off stack0) cont_args stack0 (tmp_lbl, block) <- makeFixupBlock dflags sp0 l stack2 tscp assigs return (l, tmp_lbl, stack2, block) -- (c) otherwise, the current StackMap is the StackMap for -- the continuation. But we must remember to remove any -- variables from the StackMap that are *not* live at -- the destination, because this StackMap might be used -- by fixupStack if this is a join point. | otherwise = return (l, l, stack1, []) where live = mapFindWithDefault (panic "handleBranch") l liveness stack1 = stack0 { sm_regs = filterUFM is_live (sm_regs stack0) } is_live (r,_) = r `elemRegSet` live makeFixupBlock :: DynFlags -> ByteOff -> Label -> StackMap -> CmmTickScope -> [CmmNode O O] -> UniqSM (Label, [CmmBlock]) makeFixupBlock dflags sp0 l stack tscope assigs | null assigs && sp0 == sm_sp stack = return (l, []) | otherwise = do tmp_lbl <- newBlockId let sp_off = sp0 - sm_sp stack block = blockJoin (CmmEntry tmp_lbl tscope) ( maybeAddSpAdj dflags sp0 sp_off $ blockFromList assigs ) (CmmBranch l) return (tmp_lbl, [block]) -- Sp is currently pointing to current_sp, -- we want it to point to -- (sm_sp cont_stack - sm_args cont_stack + args) -- so the difference is -- sp0 - (sm_sp cont_stack - sm_args cont_stack + args) spOffsetForCall :: ByteOff -> StackMap -> ByteOff -> ByteOff spOffsetForCall current_sp cont_stack args = current_sp - (sm_sp cont_stack - sm_args cont_stack + args) -- | create a sequence of assignments to establish the new StackMap, -- given the old StackMap. fixupStack :: StackMap -> StackMap -> [CmmNode O O] fixupStack old_stack new_stack = concatMap move new_locs where old_map = sm_regs old_stack new_locs = stackSlotRegs new_stack move (r,n) | Just (_,m) <- lookupUFM old_map r, n == m = [] | otherwise = [CmmStore (CmmStackSlot Old n) (CmmReg (CmmLocal r))] setupStackFrame :: DynFlags -> BlockId -- label of continuation -> LabelMap CmmLocalLive -- liveness -> ByteOff -- updfr -> ByteOff -- bytes of return values on stack -> StackMap -- current StackMap -> (StackMap, [CmmNode O O]) setupStackFrame dflags lbl liveness updfr_off ret_args stack0 = (cont_stack, assignments) where -- get the set of LocalRegs live in the continuation live = mapFindWithDefault Set.empty lbl liveness -- the stack from the base to updfr_off is off-limits. -- our new stack frame contains: -- * saved live variables -- * the return address [young(C) + 8] -- * the args for the call, -- which are replaced by the return values at the return -- point. -- everything up to updfr_off is off-limits -- stack1 contains updfr_off, plus everything we need to save (stack1, assignments) = allocate dflags updfr_off live stack0 -- And the Sp at the continuation is: -- sm_sp stack1 + ret_args cont_stack = stack1{ sm_sp = sm_sp stack1 + ret_args , sm_args = ret_args , sm_ret_off = updfr_off } -- ----------------------------------------------------------------------------- -- Note [diamond proc point] -- -- This special case looks for the pattern we get from a typical -- tagged case expression: -- -- Sp[young(L1)] = L1 -- if (R1 & 7) != 0 goto L1 else goto L2 -- L2: -- call [R1] returns to L1 -- L1: live: {y} -- x = R1 -- -- If we let the generic case handle this, we get -- -- Sp[-16] = L1 -- if (R1 & 7) != 0 goto L1a else goto L2 -- L2: -- Sp[-8] = y -- Sp = Sp - 16 -- call [R1] returns to L1 -- L1a: -- Sp[-8] = y -- Sp = Sp - 16 -- goto L1 -- L1: -- x = R1 -- -- The code for saving the live vars is duplicated in each branch, and -- furthermore there is an extra jump in the fast path (assuming L1 is -- a proc point, which it probably is if there is a heap check). -- -- So to fix this we want to set up the stack frame before the -- conditional jump. How do we know when to do this, and when it is -- safe? The basic idea is, when we see the assignment -- -- Sp[young(L)] = L -- -- we know that -- * we are definitely heading for L -- * there can be no more reads from another stack area, because young(L) -- overlaps with it. -- -- We don't necessarily know that everything live at L is live now -- (some might be assigned between here and the jump to L). So we -- simplify and only do the optimisation when we see -- -- (1) a block containing an assignment of a return address L -- (2) ending in a branch where one (and only) continuation goes to L, -- and no other continuations go to proc points. -- -- then we allocate the stack frame for L at the end of the block, -- before the branch. -- -- We could generalise (2), but that would make it a bit more -- complicated to handle, and this currently catches the common case. futureContinuation :: Block CmmNode O O -> Maybe BlockId futureContinuation middle = foldBlockNodesB f middle Nothing where f :: CmmNode a b -> Maybe BlockId -> Maybe BlockId f (CmmStore (CmmStackSlot (Young l) _) (CmmLit (CmmBlock _))) _ = Just l f _ r = r -- ----------------------------------------------------------------------------- -- Saving live registers -- | Given a set of live registers and a StackMap, save all the registers -- on the stack and return the new StackMap and the assignments to do -- the saving. -- allocate :: DynFlags -> ByteOff -> LocalRegSet -> StackMap -> (StackMap, [CmmNode O O]) allocate dflags ret_off live stackmap@StackMap{ sm_sp = sp0 , sm_regs = regs0 } = -- we only have to save regs that are not already in a slot let to_save = filter (not . (`elemUFM` regs0)) (Set.elems live) regs1 = filterUFM (\(r,_) -> elemRegSet r live) regs0 in -- make a map of the stack let stack = reverse $ Array.elems $ accumArray (\_ x -> x) Empty (1, toWords dflags (max sp0 ret_off)) $ ret_words ++ live_words where ret_words = [ (x, Occupied) | x <- [ 1 .. toWords dflags ret_off] ] live_words = [ (toWords dflags x, Occupied) | (r,off) <- nonDetEltsUFM regs1, -- See Note [Unique Determinism and code generation] let w = localRegBytes dflags r, x <- [ off, off - wORD_SIZE dflags .. off - w + 1] ] in -- Pass over the stack: find slots to save all the new live variables, -- choosing the oldest slots first (hence a foldr). let save slot ([], stack, n, assigs, regs) -- no more regs to save = ([], slot:stack, plusW dflags n 1, assigs, regs) save slot (to_save, stack, n, assigs, regs) = case slot of Occupied -> (to_save, Occupied:stack, plusW dflags n 1, assigs, regs) Empty | Just (stack', r, to_save') <- select_save to_save (slot:stack) -> let assig = CmmStore (CmmStackSlot Old n') (CmmReg (CmmLocal r)) n' = plusW dflags n 1 in (to_save', stack', n', assig : assigs, (r,(r,n')):regs) | otherwise -> (to_save, slot:stack, plusW dflags n 1, assigs, regs) -- we should do better here: right now we'll fit the smallest first, -- but it would make more sense to fit the biggest first. select_save :: [LocalReg] -> [StackSlot] -> Maybe ([StackSlot], LocalReg, [LocalReg]) select_save regs stack = go regs [] where go [] _no_fit = Nothing go (r:rs) no_fit | Just rest <- dropEmpty words stack = Just (replicate words Occupied ++ rest, r, rs++no_fit) | otherwise = go rs (r:no_fit) where words = localRegWords dflags r -- fill in empty slots as much as possible (still_to_save, save_stack, n, save_assigs, save_regs) = foldr save (to_save, [], 0, [], []) stack -- push any remaining live vars on the stack (push_sp, push_assigs, push_regs) = foldr push (n, [], []) still_to_save where push r (n, assigs, regs) = (n', assig : assigs, (r,(r,n')) : regs) where n' = n + localRegBytes dflags r assig = CmmStore (CmmStackSlot Old n') (CmmReg (CmmLocal r)) trim_sp | not (null push_regs) = push_sp | otherwise = plusW dflags n (- length (takeWhile isEmpty save_stack)) final_regs = regs1 `addListToUFM` push_regs `addListToUFM` save_regs in -- XXX should be an assert if ( n /= max sp0 ret_off ) then pprPanic "allocate" (ppr n <+> ppr sp0 <+> ppr ret_off) else if (trim_sp .&. (wORD_SIZE dflags - 1)) /= 0 then pprPanic "allocate2" (ppr trim_sp <+> ppr final_regs <+> ppr push_sp) else ( stackmap { sm_regs = final_regs , sm_sp = trim_sp } , push_assigs ++ save_assigs ) -- ----------------------------------------------------------------------------- -- Manifesting Sp -- | Manifest Sp: turn all the CmmStackSlots into CmmLoads from Sp. The -- block looks like this: -- -- middle_pre -- the middle nodes -- Sp = Sp + sp_off -- Sp adjustment goes here -- last -- the last node -- -- And we have some extra blocks too (that don't contain Sp adjustments) -- -- The adjustment for middle_pre will be different from that for -- middle_post, because the Sp adjustment intervenes. -- manifestSp :: DynFlags -> LabelMap StackMap -- StackMaps for other blocks -> StackMap -- StackMap for this block -> ByteOff -- Sp on entry to the block -> ByteOff -- SpHigh -> CmmNode C O -- first node -> [CmmNode O O] -- middle -> ByteOff -- sp_off -> CmmNode O C -- last node -> [CmmBlock] -- new blocks -> [CmmBlock] -- final blocks with Sp manifest manifestSp dflags stackmaps stack0 sp0 sp_high first middle_pre sp_off last fixup_blocks = final_block : fixup_blocks' where area_off = getAreaOff stackmaps adj_pre_sp, adj_post_sp :: CmmNode e x -> CmmNode e x adj_pre_sp = mapExpDeep (areaToSp dflags sp0 sp_high area_off) adj_post_sp = mapExpDeep (areaToSp dflags (sp0 - sp_off) sp_high area_off) final_middle = maybeAddSpAdj dflags sp0 sp_off . blockFromList . map adj_pre_sp . elimStackStores stack0 stackmaps area_off $ middle_pre final_last = optStackCheck (adj_post_sp last) final_block = blockJoin first final_middle final_last fixup_blocks' = map (mapBlock3' (id, adj_post_sp, id)) fixup_blocks getAreaOff :: LabelMap StackMap -> (Area -> StackLoc) getAreaOff _ Old = 0 getAreaOff stackmaps (Young l) = case mapLookup l stackmaps of Just sm -> sm_sp sm - sm_args sm Nothing -> pprPanic "getAreaOff" (ppr l) maybeAddSpAdj :: DynFlags -> ByteOff -> ByteOff -> Block CmmNode O O -> Block CmmNode O O maybeAddSpAdj dflags sp0 sp_off block = add_initial_unwind $ add_adj_unwind $ adj block where adj block | sp_off /= 0 = block `blockSnoc` CmmAssign spReg (cmmOffset dflags spExpr sp_off) | otherwise = block -- Add unwind pseudo-instruction at the beginning of each block to -- document Sp level for debugging add_initial_unwind block | debugLevel dflags > 0 = CmmUnwind [(Sp, Just sp_unwind)] `blockCons` block | otherwise = block where sp_unwind = CmmRegOff spReg (sp0 - wORD_SIZE dflags) -- Add unwind pseudo-instruction right after the Sp adjustment -- if there is one. add_adj_unwind block | debugLevel dflags > 0 , sp_off /= 0 = block `blockSnoc` CmmUnwind [(Sp, Just sp_unwind)] | otherwise = block where sp_unwind = CmmRegOff spReg (sp0 - wORD_SIZE dflags - sp_off) {- Note [SP old/young offsets] Sp(L) is the Sp offset on entry to block L relative to the base of the OLD area. SpArgs(L) is the size of the young area for L, i.e. the number of arguments. - in block L, each reference to [old + N] turns into [Sp + Sp(L) - N] - in block L, each reference to [young(L') + N] turns into [Sp + Sp(L) - Sp(L') + SpArgs(L') - N] - be careful with the last node of each block: Sp has already been adjusted to be Sp + Sp(L) - Sp(L') -} areaToSp :: DynFlags -> ByteOff -> ByteOff -> (Area -> StackLoc) -> CmmExpr -> CmmExpr areaToSp dflags sp_old _sp_hwm area_off (CmmStackSlot area n) = cmmOffset dflags spExpr (sp_old - area_off area - n) -- Replace (CmmStackSlot area n) with an offset from Sp areaToSp dflags _ sp_hwm _ (CmmLit CmmHighStackMark) = mkIntExpr dflags sp_hwm -- Replace CmmHighStackMark with the number of bytes of stack used, -- the sp_hwm. See Note [Stack usage] in GHC.StgToCmm.Heap areaToSp dflags _ _ _ (CmmMachOp (MO_U_Lt _) args) | falseStackCheck args = zeroExpr dflags areaToSp dflags _ _ _ (CmmMachOp (MO_U_Ge _) args) | falseStackCheck args = mkIntExpr dflags 1 -- Replace a stack-overflow test that cannot fail with a no-op -- See Note [Always false stack check] areaToSp _ _ _ _ other = other -- | Determine whether a stack check cannot fail. falseStackCheck :: [CmmExpr] -> Bool falseStackCheck [ CmmMachOp (MO_Sub _) [ CmmRegOff (CmmGlobal Sp) x_off , CmmLit (CmmInt y_lit _)] , CmmReg (CmmGlobal SpLim)] = fromIntegral x_off >= y_lit falseStackCheck _ = False -- Note [Always false stack check] -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- We can optimise stack checks of the form -- -- if ((Sp + x) - y < SpLim) then .. else .. -- -- where are non-negative integer byte offsets. Since we know that -- SpLim <= Sp (remember the stack grows downwards), this test must -- yield False if (x >= y), so we can rewrite the comparison to False. -- A subsequent sinking pass will later drop the dead code. -- Optimising this away depends on knowing that SpLim <= Sp, so it is -- really the job of the stack layout algorithm, hence we do it now. -- -- The control flow optimiser may negate a conditional to increase -- the likelihood of a fallthrough if the branch is not taken. But -- not every conditional is inverted as the control flow optimiser -- places some requirements on the predecessors of both branch targets. -- So we better look for the inverted comparison too. optStackCheck :: CmmNode O C -> CmmNode O C optStackCheck n = -- Note [Always false stack check] case n of CmmCondBranch (CmmLit (CmmInt 0 _)) _true false _ -> CmmBranch false CmmCondBranch (CmmLit (CmmInt _ _)) true _false _ -> CmmBranch true other -> other -- ----------------------------------------------------------------------------- -- | Eliminate stores of the form -- -- Sp[area+n] = r -- -- when we know that r is already in the same slot as Sp[area+n]. We -- could do this in a later optimisation pass, but that would involve -- a separate analysis and we already have the information to hand -- here. It helps clean up some extra stack stores in common cases. -- -- Note that we may have to modify the StackMap as we walk through the -- code using procMiddle, since an assignment to a variable in the -- StackMap will invalidate its mapping there. -- elimStackStores :: StackMap -> LabelMap StackMap -> (Area -> ByteOff) -> [CmmNode O O] -> [CmmNode O O] elimStackStores stackmap stackmaps area_off nodes = go stackmap nodes where go _stackmap [] = [] go stackmap (n:ns) = case n of CmmStore (CmmStackSlot area m) (CmmReg (CmmLocal r)) | Just (_,off) <- lookupUFM (sm_regs stackmap) r , area_off area + m == off -> go stackmap ns _otherwise -> n : go (procMiddle stackmaps n stackmap) ns -- ----------------------------------------------------------------------------- -- Update info tables to include stack liveness setInfoTableStackMap :: DynFlags -> LabelMap StackMap -> CmmDecl -> CmmDecl setInfoTableStackMap dflags stackmaps (CmmProc top_info@TopInfo{..} l v g) = CmmProc top_info{ info_tbls = mapMapWithKey fix_info info_tbls } l v g where fix_info lbl info_tbl@CmmInfoTable{ cit_rep = StackRep _ } = info_tbl { cit_rep = StackRep (get_liveness lbl) } fix_info _ other = other get_liveness :: BlockId -> Liveness get_liveness lbl = case mapLookup lbl stackmaps of Nothing -> pprPanic "setInfoTableStackMap" (ppr lbl <+> ppr info_tbls) Just sm -> stackMapToLiveness dflags sm setInfoTableStackMap _ _ d = d stackMapToLiveness :: DynFlags -> StackMap -> Liveness stackMapToLiveness dflags StackMap{..} = reverse $ Array.elems $ accumArray (\_ x -> x) True (toWords dflags sm_ret_off + 1, toWords dflags (sm_sp - sm_args)) live_words where live_words = [ (toWords dflags off, False) | (r,off) <- nonDetEltsUFM sm_regs , isGcPtrType (localRegType r) ] -- See Note [Unique Determinism and code generation] -- ----------------------------------------------------------------------------- -- Pass 2 -- ----------------------------------------------------------------------------- insertReloadsAsNeeded :: DynFlags -> ProcPointSet -> LabelMap StackMap -> BlockId -> [CmmBlock] -> UniqSM [CmmBlock] insertReloadsAsNeeded dflags procpoints final_stackmaps entry blocks = do toBlockList . fst <$> rewriteCmmBwd liveLattice rewriteCC (ofBlockList entry blocks) mapEmpty where rewriteCC :: RewriteFun CmmLocalLive rewriteCC (BlockCC e_node middle0 x_node) fact_base0 = do let entry_label = entryLabel e_node stackmap = case mapLookup entry_label final_stackmaps of Just sm -> sm Nothing -> panic "insertReloadsAsNeeded: rewriteCC: stackmap" -- Merge the liveness from successor blocks and analyse the last -- node. joined = gen_kill dflags x_node $! joinOutFacts liveLattice x_node fact_base0 -- What is live at the start of middle0. live_at_middle0 = foldNodesBwdOO (gen_kill dflags) middle0 joined -- If this is a procpoint we need to add the reloads, but only if -- they're actually live. Furthermore, nothing is live at the entry -- to a proc point. (middle1, live_with_reloads) | entry_label `setMember` procpoints = let reloads = insertReloads dflags stackmap live_at_middle0 in (foldr blockCons middle0 reloads, emptyRegSet) | otherwise = (middle0, live_at_middle0) -- Final liveness for this block. !fact_base2 = mapSingleton entry_label live_with_reloads return (BlockCC e_node middle1 x_node, fact_base2) insertReloads :: DynFlags -> StackMap -> CmmLocalLive -> [CmmNode O O] insertReloads dflags stackmap live = [ CmmAssign (CmmLocal reg) -- This cmmOffset basically corresponds to manifesting -- @CmmStackSlot Old sp_off@, see Note [SP old/young offsets] (CmmLoad (cmmOffset dflags spExpr (sp_off - reg_off)) (localRegType reg)) | (reg, reg_off) <- stackSlotRegs stackmap , reg `elemRegSet` live ] where sp_off = sm_sp stackmap -- ----------------------------------------------------------------------------- -- Lowering safe foreign calls {- Note [Lower safe foreign calls] We start with Sp[young(L1)] = L1 ,----------------------- | r1 = foo(x,y,z) returns to L1 '----------------------- L1: R1 = r1 -- copyIn, inserted by mkSafeCall ... the stack layout algorithm will arrange to save and reload everything live across the call. Our job now is to expand the call so we get Sp[young(L1)] = L1 ,----------------------- | SAVE_THREAD_STATE() | token = suspendThread(BaseReg, interruptible) | r = foo(x,y,z) | BaseReg = resumeThread(token) | LOAD_THREAD_STATE() | R1 = r -- copyOut | jump Sp[0] '----------------------- L1: r = R1 -- copyIn, inserted by mkSafeCall ... Note the copyOut, which saves the results in the places that L1 is expecting them (see Note [safe foreign call convention]). Note also that safe foreign call is replace by an unsafe one in the Cmm graph. -} lowerSafeForeignCall :: DynFlags -> CmmBlock -> UniqSM CmmBlock lowerSafeForeignCall dflags block | (entry@(CmmEntry _ tscp), middle, CmmForeignCall { .. }) <- blockSplit block = do -- Both 'id' and 'new_base' are KindNonPtr because they're -- RTS-only objects and are not subject to garbage collection id <- newTemp (bWord dflags) new_base <- newTemp (cmmRegType dflags baseReg) let (caller_save, caller_load) = callerSaveVolatileRegs dflags save_state_code <- saveThreadState dflags load_state_code <- loadThreadState dflags let suspend = save_state_code <*> caller_save <*> mkMiddle (callSuspendThread dflags id intrbl) midCall = mkUnsafeCall tgt res args resume = mkMiddle (callResumeThread new_base id) <*> -- Assign the result to BaseReg: we -- might now have a different Capability! mkAssign baseReg (CmmReg (CmmLocal new_base)) <*> caller_load <*> load_state_code (_, regs, copyout) = copyOutOflow dflags NativeReturn Jump (Young succ) (map (CmmReg . CmmLocal) res) ret_off [] -- NB. after resumeThread returns, the top-of-stack probably contains -- the stack frame for succ, but it might not: if the current thread -- received an exception during the call, then the stack might be -- different. Hence we continue by jumping to the top stack frame, -- not by jumping to succ. jump = CmmCall { cml_target = entryCode dflags $ CmmLoad spExpr (bWord dflags) , cml_cont = Just succ , cml_args_regs = regs , cml_args = widthInBytes (wordWidth dflags) , cml_ret_args = ret_args , cml_ret_off = ret_off } graph' <- lgraphOfAGraph ( suspend <*> midCall <*> resume <*> copyout <*> mkLast jump, tscp) case toBlockList graph' of [one] -> let (_, middle', last) = blockSplit one in return (blockJoin entry (middle `blockAppend` middle') last) _ -> panic "lowerSafeForeignCall0" -- Block doesn't end in a safe foreign call: | otherwise = return block foreignLbl :: FastString -> CmmExpr foreignLbl name = CmmLit (CmmLabel (mkForeignLabel name Nothing ForeignLabelInExternalPackage IsFunction)) -- void * suspendThread (StgRegTable *, bool interruptible); callSuspendThread :: DynFlags -> LocalReg -> Bool -> CmmNode O O callSuspendThread dflags id intrbl = CmmUnsafeForeignCall (ForeignTarget (foreignLbl (fsLit "suspendThread")) (ForeignConvention CCallConv [AddrHint, NoHint] [AddrHint] CmmMayReturn AddrRep [AddrRep, Word32Rep])) [id] [baseExpr, mkIntExpr dflags (fromEnum intrbl)] -- StgRegTable * resumeThread (void *); callResumeThread :: LocalReg -> LocalReg -> CmmNode O O callResumeThread new_base id = CmmUnsafeForeignCall (ForeignTarget (foreignLbl (fsLit "resumeThread")) (ForeignConvention CCallConv [AddrHint] [AddrHint] CmmMayReturn AddrRep [AddrRep])) [new_base] [CmmReg (CmmLocal id)] -- ----------------------------------------------------------------------------- plusW :: DynFlags -> ByteOff -> WordOff -> ByteOff plusW dflags b w = b + w * wORD_SIZE dflags data StackSlot = Occupied | Empty -- Occupied: a return address or part of an update frame instance Outputable StackSlot where ppr Occupied = text "XXX" ppr Empty = text "---" dropEmpty :: WordOff -> [StackSlot] -> Maybe [StackSlot] dropEmpty 0 ss = Just ss dropEmpty n (Empty : ss) = dropEmpty (n-1) ss dropEmpty _ _ = Nothing isEmpty :: StackSlot -> Bool isEmpty Empty = True isEmpty _ = False localRegBytes :: DynFlags -> LocalReg -> ByteOff localRegBytes dflags r = roundUpToWords dflags (widthInBytes (typeWidth (localRegType r))) localRegWords :: DynFlags -> LocalReg -> WordOff localRegWords dflags = toWords dflags . localRegBytes dflags toWords :: DynFlags -> ByteOff -> WordOff toWords dflags x = x `quot` wORD_SIZE dflags stackSlotRegs :: StackMap -> [(LocalReg, StackLoc)] stackSlotRegs sm = nonDetEltsUFM (sm_regs sm) -- See Note [Unique Determinism and code generation]