-%\r
-% (c) The AQUA Project, Glasgow University, 1994-1998\r
-%\r
-\section[CoreUnfold]{Core-syntax unfoldings}\r
-\r
-Unfoldings (which can travel across module boundaries) are in Core\r
-syntax (namely @CoreExpr@s).\r
-\r
-The type @Unfolding@ sits ``above'' simply-Core-expressions\r
-unfoldings, capturing ``higher-level'' things we know about a binding,\r
-usually things that the simplifier found out (e.g., ``it's a\r
-literal''). In the corner of a @CoreUnfolding@ unfolding, you will\r
-find, unsurprisingly, a Core expression.\r
-\r
-\begin{code}\r
-module CoreUnfold (\r
- Unfolding(..), UnfoldingGuidance, -- types\r
-\r
- noUnfolding, mkUnfolding, getUnfoldingTemplate,\r
- isEvaldUnfolding, hasUnfolding,\r
-\r
- couldBeSmallEnoughToInline, \r
- certainlySmallEnoughToInline, \r
- okToUnfoldInHiFile,\r
-\r
- calcUnfoldingGuidance,\r
-\r
- callSiteInline, blackListed\r
- ) where\r
-\r
-#include "HsVersions.h"\r
-\r
-import CmdLineOpts ( opt_UF_CreationThreshold,\r
- opt_UF_UseThreshold,\r
- opt_UF_ScrutConDiscount,\r
- opt_UF_FunAppDiscount,\r
- opt_UF_PrimArgDiscount,\r
- opt_UF_KeenessFactor,\r
- opt_UF_CheapOp, opt_UF_DearOp, opt_UF_NoRepLit,\r
- opt_UnfoldCasms, opt_PprStyle_Debug,\r
- opt_D_dump_inlinings\r
- )\r
-import CoreSyn\r
-import PprCore ( pprCoreExpr )\r
-import OccurAnal ( occurAnalyseGlobalExpr )\r
-import BinderInfo ( )\r
-import CoreUtils ( coreExprType, exprIsTrivial, mkFormSummary, whnfOrBottom,\r
- FormSummary(..) )\r
-import Id ( Id, idType, idUnique, isId, \r
- getIdSpecialisation, getInlinePragma, getIdUnfolding\r
- )\r
-import VarSet\r
-import Const ( Con(..), isLitLitLit, isWHNFCon )\r
-import PrimOp ( PrimOp(..), primOpIsDupable )\r
-import IdInfo ( ArityInfo(..), InlinePragInfo(..), OccInfo(..) )\r
-import TyCon ( tyConFamilySize )\r
-import Type ( splitAlgTyConApp_maybe, splitFunTy_maybe )\r
-import Const ( isNoRepLit )\r
-import Unique ( Unique, buildIdKey, augmentIdKey, runSTRepIdKey )\r
-import Maybes ( maybeToBool )\r
-import Bag\r
-import Util ( isIn, lengthExceeds )\r
-import Outputable\r
-\end{code}\r
-\r
-%************************************************************************\r
-%* *\r
-\subsection{@Unfolding@ and @UnfoldingGuidance@ types}\r
-%* *\r
-%************************************************************************\r
-\r
-\begin{code}\r
-data Unfolding\r
- = NoUnfolding\r
-\r
- | OtherCon [Con] -- It ain't one of these\r
- -- (OtherCon xs) also indicates that something has been evaluated\r
- -- and hence there's no point in re-evaluating it.\r
- -- OtherCon [] is used even for non-data-type values\r
- -- to indicated evaluated-ness. Notably:\r
- -- data C = C !(Int -> Int)\r
- -- case x of { C f -> ... }\r
- -- Here, f gets an OtherCon [] unfolding.\r
-\r
- | CoreUnfolding -- An unfolding with redundant cached information\r
- FormSummary -- Tells whether the template is a WHNF or bottom\r
- UnfoldingGuidance -- Tells about the *size* of the template.\r
- CoreExpr -- Template; binder-info is correct\r
-\end{code}\r
-\r
-\begin{code}\r
-noUnfolding = NoUnfolding\r
-\r
-mkUnfolding expr\r
- = let\r
- -- strictness mangling (depends on there being no CSE)\r
- ufg = calcUnfoldingGuidance opt_UF_CreationThreshold expr\r
- occ = occurAnalyseGlobalExpr expr\r
- in\r
- CoreUnfolding (mkFormSummary expr) ufg occ\r
-\r
-getUnfoldingTemplate :: Unfolding -> CoreExpr\r
-getUnfoldingTemplate (CoreUnfolding _ _ expr) = expr\r
-getUnfoldingTemplate other = panic "getUnfoldingTemplate"\r
-\r
-isEvaldUnfolding :: Unfolding -> Bool\r
-isEvaldUnfolding (OtherCon _) = True\r
-isEvaldUnfolding (CoreUnfolding ValueForm _ expr) = True\r
-isEvaldUnfolding other = False\r
-\r
-hasUnfolding :: Unfolding -> Bool\r
-hasUnfolding NoUnfolding = False\r
-hasUnfolding other = True\r
-\r
-data UnfoldingGuidance\r
- = UnfoldNever\r
- | UnfoldAlways -- There is no "original" definition,\r
- -- so you'd better unfold. Or: something\r
- -- so cheap to unfold (e.g., 1#) that\r
- -- you should do it absolutely always.\r
-\r
- | UnfoldIfGoodArgs Int -- and "n" value args\r
-\r
- [Int] -- Discount if the argument is evaluated.\r
- -- (i.e., a simplification will definitely\r
- -- be possible). One elt of the list per *value* arg.\r
-\r
- Int -- The "size" of the unfolding; to be elaborated\r
- -- later. ToDo\r
-\r
- Int -- Scrutinee discount: the discount to substract if the thing is in\r
- -- a context (case (thing args) of ...),\r
- -- (where there are the right number of arguments.)\r
-\end{code}\r
-\r
-\begin{code}\r
-instance Outputable UnfoldingGuidance where\r
- ppr UnfoldAlways = ptext SLIT("ALWAYS")\r
- ppr UnfoldNever = ptext SLIT("NEVER")\r
- ppr (UnfoldIfGoodArgs v cs size discount)\r
- = hsep [ptext SLIT("IF_ARGS"), int v,\r
- if null cs -- always print *something*\r
- then char 'X'\r
- else hcat (map (text . show) cs),\r
- int size,\r
- int discount ]\r
-\end{code}\r
-\r
-\r
-%************************************************************************\r
-%* *\r
-\subsection[calcUnfoldingGuidance]{Calculate ``unfolding guidance'' for an expression}\r
-%* *\r
-%************************************************************************\r
-\r
-\begin{code}\r
-calcUnfoldingGuidance\r
- :: Int -- bomb out if size gets bigger than this\r
- -> CoreExpr -- expression to look at\r
- -> UnfoldingGuidance\r
-calcUnfoldingGuidance bOMB_OUT_SIZE expr\r
- | exprIsTrivial expr -- Often trivial expressions are never bound\r
- -- to an expression, but it can happen. For\r
- -- example, the Id for a nullary constructor has\r
- -- a trivial expression as its unfolding, and\r
- -- we want to make sure that we always unfold it.\r
- = UnfoldAlways\r
- \r
- | otherwise\r
- = case collectBinders expr of { (binders, body) ->\r
- let\r
- val_binders = filter isId binders\r
- in\r
- case (sizeExpr bOMB_OUT_SIZE val_binders body) of\r
-\r
- TooBig -> UnfoldNever\r
-\r
- SizeIs size cased_args scrut_discount\r
- -> UnfoldIfGoodArgs\r
- (length val_binders)\r
- (map discount_for val_binders)\r
- (I# size)\r
- (I# scrut_discount)\r
- where \r
- discount_for b \r
- | num_cases == 0 = 0\r
- | is_fun_ty = num_cases * opt_UF_FunAppDiscount\r
- | is_data_ty = num_cases * tyConFamilySize tycon * opt_UF_ScrutConDiscount\r
- | otherwise = num_cases * opt_UF_PrimArgDiscount\r
- where\r
- num_cases = foldlBag (\n b' -> if b==b' then n+1 else n) 0 cased_args\r
- -- Count occurrences of b in cased_args\r
- arg_ty = idType b\r
- is_fun_ty = maybeToBool (splitFunTy_maybe arg_ty)\r
- (is_data_ty, tycon) = case (splitAlgTyConApp_maybe (idType b)) of\r
- Nothing -> (False, panic "discount")\r
- Just (tc,_,_) -> (True, tc)\r
- }\r
-\end{code}\r
-\r
-\begin{code}\r
-sizeExpr :: Int -- Bomb out if it gets bigger than this\r
- -> [Id] -- Arguments; we're interested in which of these\r
- -- get case'd\r
- -> CoreExpr\r
- -> ExprSize\r
-\r
-sizeExpr (I# bOMB_OUT_SIZE) args expr\r
- = size_up expr\r
- where\r
- size_up (Type t) = sizeZero -- Types cost nothing\r
- size_up (Var v) = sizeOne\r
-\r
- size_up (Note InlineMe _) = sizeTwo -- The idea is that this is one more\r
- -- than the size of the "call" (i.e. 1)\r
- -- We want to reply "no" to noSizeIncrease\r
- -- for a bare reference (i.e. applied to no args) \r
- -- to an INLINE thing\r
-\r
- size_up (Note _ body) = size_up body -- Notes cost nothing\r
-\r
- size_up (App fun (Type t)) = size_up fun\r
- size_up (App fun arg) = size_up_app fun `addSize` size_up arg\r
-\r
- size_up (Con con args) = foldr (addSize . size_up) \r
- (size_up_con con args)\r
- args\r
-\r
- size_up (Lam b e) | isId b = size_up e `addSizeN` 1\r
- | otherwise = size_up e\r
-\r
- size_up (Let (NonRec binder rhs) body)\r
- = nukeScrutDiscount (size_up rhs) `addSize`\r
- size_up body `addSizeN`\r
- 1 -- For the allocation\r
-\r
- size_up (Let (Rec pairs) body)\r
- = nukeScrutDiscount rhs_size `addSize`\r
- size_up body `addSizeN`\r
- length pairs -- For the allocation\r
- where\r
- rhs_size = foldr (addSize . size_up . snd) sizeZero pairs\r
-\r
- size_up (Case scrut _ alts)\r
- = nukeScrutDiscount (size_up scrut) `addSize`\r
- arg_discount scrut `addSize`\r
- foldr (addSize . size_up_alt) sizeZero alts `addSizeN`\r
- case (splitAlgTyConApp_maybe (coreExprType scrut)) of\r
- Nothing -> 1\r
- Just (tc,_,_) -> tyConFamilySize tc\r
-\r
- ------------ \r
- -- A function application with at least one value argument\r
- -- so if the function is an argument give it an arg-discount\r
- size_up_app (App fun arg) = size_up_app fun `addSize` size_up arg\r
- size_up_app fun = arg_discount fun `addSize` size_up fun\r
-\r
- ------------ \r
- size_up_alt (con, bndrs, rhs) = size_up rhs\r
- -- Don't charge for args, so that wrappers look cheap\r
-\r
- ------------\r
- size_up_con (Literal lit) args | isNoRepLit lit = sizeN opt_UF_NoRepLit\r
- | otherwise = sizeOne\r
-\r
- size_up_con (DataCon dc) args = conSizeN (valArgCount args)\r
- \r
- size_up_con (PrimOp op) args = foldr addSize (sizeN op_cost) (map arg_discount args)\r
- -- Give an arg-discount if a primop is applies to\r
- -- one of the function's arguments\r
- where\r
- op_cost | primOpIsDupable op = opt_UF_CheapOp\r
- | otherwise = opt_UF_DearOp\r
-\r
- ------------\r
- -- We want to record if we're case'ing, or applying, an argument\r
- arg_discount (Var v) | v `is_elem` args = scrutArg v\r
- arg_discount other = sizeZero\r
-\r
- is_elem :: Id -> [Id] -> Bool\r
- is_elem = isIn "size_up_scrut"\r
-\r
- ------------\r
- -- These addSize things have to be here because\r
- -- I don't want to give them bOMB_OUT_SIZE as an argument\r
-\r
- addSizeN TooBig _ = TooBig\r
- addSizeN (SizeIs n xs d) (I# m)\r
- | n_tot -# d <# bOMB_OUT_SIZE = SizeIs n_tot xs d\r
- | otherwise = TooBig\r
- where\r
- n_tot = n +# m\r
- \r
- addSize TooBig _ = TooBig\r
- addSize _ TooBig = TooBig\r
- addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)\r
- | (n_tot -# d_tot) <# bOMB_OUT_SIZE = SizeIs n_tot xys d_tot\r
- | otherwise = TooBig\r
- where\r
- n_tot = n1 +# n2\r
- d_tot = d1 +# d2\r
- xys = xs `unionBags` ys\r
-\end{code}\r
-\r
-Code for manipulating sizes\r
-\r
-\begin{code}\r
-\r
-data ExprSize = TooBig\r
- | SizeIs Int# -- Size found\r
- (Bag Id) -- Arguments cased herein\r
- Int# -- Size to subtract if result is scrutinised \r
- -- by a case expression\r
-\r
-sizeZero = SizeIs 0# emptyBag 0#\r
-sizeOne = SizeIs 1# emptyBag 0#\r
-sizeTwo = SizeIs 2# emptyBag 0#\r
-sizeN (I# n) = SizeIs n emptyBag 0#\r
-conSizeN (I# n) = SizeIs 1# emptyBag (n +# 1#)\r
- -- Treat constructors as size 1, that unfoldAlways responsds 'False'\r
- -- when asked about 'x' when x is bound to (C 3#).\r
- -- This avoids gratuitous 'ticks' when x itself appears as an\r
- -- atomic constructor argument.\r
- \r
-scrutArg v = SizeIs 0# (unitBag v) 0#\r
-\r
-nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#\r
-nukeScrutDiscount TooBig = TooBig\r
-\end{code}\r
-\r
-\r
-%************************************************************************\r
-%* *\r
-\subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}\r
-%* *\r
-%************************************************************************\r
-\r
-We have very limited information about an unfolding expression: (1)~so\r
-many type arguments and so many value arguments expected---for our\r
-purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''\r
-a single integer. (3)~An ``argument info'' vector. For this, what we\r
-have at the moment is a Boolean per argument position that says, ``I\r
-will look with great favour on an explicit constructor in this\r
-position.'' (4)~The ``discount'' to subtract if the expression\r
-is being scrutinised. \r
-\r
-Assuming we have enough type- and value arguments (if not, we give up\r
-immediately), then we see if the ``discounted size'' is below some\r
-(semi-arbitrary) threshold. It works like this: for every argument\r
-position where we're looking for a constructor AND WE HAVE ONE in our\r
-hands, we get a (again, semi-arbitrary) discount [proportion to the\r
-number of constructors in the type being scrutinized].\r
-\r
-If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})\r
-and the expression in question will evaluate to a constructor, we use\r
-the computed discount size *for the result only* rather than\r
-computing the argument discounts. Since we know the result of\r
-the expression is going to be taken apart, discounting its size\r
-is more accurate (see @sizeExpr@ above for how this discount size\r
-is computed).\r
-\r
-We use this one to avoid exporting inlinings that we ``couldn't possibly\r
-use'' on the other side. Can be overridden w/ flaggery.\r
-Just the same as smallEnoughToInline, except that it has no actual arguments.\r
-\r
-\begin{code}\r
-couldBeSmallEnoughToInline :: UnfoldingGuidance -> Bool\r
-couldBeSmallEnoughToInline UnfoldNever = False\r
-couldBeSmallEnoughToInline other = True\r
-\r
-certainlySmallEnoughToInline :: UnfoldingGuidance -> Bool\r
-certainlySmallEnoughToInline UnfoldNever = False\r
-certainlySmallEnoughToInline UnfoldAlways = True\r
-certainlySmallEnoughToInline (UnfoldIfGoodArgs _ _ size _) = size <= opt_UF_UseThreshold\r
-\end{code}\r
-\r
-@okToUnfoldInHifile@ is used when emitting unfolding info into an interface\r
-file to determine whether an unfolding candidate really should be unfolded.\r
-The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted\r
-into interface files. \r
-\r
-The reason for inlining expressions containing _casm_s into interface files\r
-is that these fragments of C are likely to mention functions/#defines that\r
-will be out-of-scope when inlined into another module. This is not an\r
-unfixable problem for the user (just need to -#include the approp. header\r
-file), but turning it off seems to the simplest thing to do.\r
-\r
-\begin{code}\r
-okToUnfoldInHiFile :: CoreExpr -> Bool\r
-okToUnfoldInHiFile e = opt_UnfoldCasms || go e\r
- where\r
- -- Race over an expression looking for CCalls..\r
- go (Var _) = True\r
- go (Con (Literal lit) _) = not (isLitLitLit lit)\r
- go (Con (PrimOp op) args) = okToUnfoldPrimOp op && all go args\r
- go (Con con args) = True -- con args are always atomic\r
- go (App fun arg) = go fun && go arg\r
- go (Lam _ body) = go body\r
- go (Let binds body) = and (map go (body :rhssOfBind binds))\r
- go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts))\r
- go (Note _ body) = go body\r
- go (Type _) = True\r
-\r
- -- ok to unfold a PrimOp as long as it's not a _casm_\r
- okToUnfoldPrimOp (CCallOp _ is_casm _ _) = not is_casm\r
- okToUnfoldPrimOp _ = True\r
-\end{code}\r
-\r
-\r
-%************************************************************************\r
-%* *\r
-\subsection{callSiteInline}\r
-%* *\r
-%************************************************************************\r
-\r
-This is the key function. It decides whether to inline a variable at a call site\r
-\r
-callSiteInline is used at call sites, so it is a bit more generous.\r
-It's a very important function that embodies lots of heuristics.\r
-A non-WHNF can be inlined if it doesn't occur inside a lambda,\r
-and occurs exactly once or \r
- occurs once in each branch of a case and is small\r
-\r
-If the thing is in WHNF, there's no danger of duplicating work, \r
-so we can inline if it occurs once, or is small\r
-\r
-\begin{code}\r
-callSiteInline :: Bool -- True <=> the Id is black listed\r
- -> Bool -- 'inline' note at call site\r
- -> Id -- The Id\r
- -> [CoreExpr] -- Arguments\r
- -> Bool -- True <=> continuation is interesting\r
- -> Maybe CoreExpr -- Unfolding, if any\r
-\r
-\r
-callSiteInline black_listed inline_call id args interesting_cont\r
- = case getIdUnfolding id of {\r
- NoUnfolding -> Nothing ;\r
- OtherCon _ -> Nothing ;\r
- CoreUnfolding form guidance unf_template ->\r
-\r
- let\r
- result | yes_or_no = Just unf_template\r
- | otherwise = Nothing\r
-\r
- inline_prag = getInlinePragma id\r
- arg_infos = map interestingArg val_args\r
- val_args = filter isValArg args\r
- whnf = whnfOrBottom form\r
-\r
- yes_or_no =\r
- case inline_prag of\r
- IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False\r
- IMustNotBeINLINEd -> False\r
- IAmALoopBreaker -> False\r
- IMustBeINLINEd -> True -- Overrides absolutely everything, including the black list\r
- ICanSafelyBeINLINEd in_lam one_br -> consider in_lam one_br\r
- NoInlinePragInfo -> consider InsideLam False\r
-\r
- consider in_lam one_branch \r
- | black_listed = False\r
- | inline_call = True\r
- | one_branch -- Be very keen to inline something if this is its unique occurrence; that\r
- -- gives a good chance of eliminating the original binding for the thing.\r
- -- The only time we hold back is when substituting inside a lambda;\r
- -- then if the context is totally uninteresting (not applied, not scrutinised)\r
- -- there is no point in substituting because it might just increase allocation.\r
- = case in_lam of\r
- NotInsideLam -> True\r
- InsideLam -> whnf && (not (null args) || interesting_cont)\r
-\r
- | otherwise -- Occurs (textually) more than once, so look at its size\r
- = case guidance of\r
- UnfoldAlways -> True\r
- UnfoldNever -> False\r
- UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount\r
- | enough_args && size <= (n_vals_wanted + 1)\r
- -- No size increase\r
- -- Size of call is n_vals_wanted (+1 for the function)\r
- -> case in_lam of\r
- NotInsideLam -> True\r
- InsideLam -> whnf\r
-\r
- | not (or arg_infos || really_interesting_cont)\r
- -- If it occurs more than once, there must be something interesting \r
- -- about some argument, or the result, to make it worth inlining\r
- -> False\r
- \r
- | otherwise\r
- -> case in_lam of\r
- NotInsideLam -> small_enough\r
- InsideLam -> whnf && small_enough\r
-\r
- where\r
- n_args = length arg_infos\r
- enough_args = n_args >= n_vals_wanted\r
- really_interesting_cont | n_args < n_vals_wanted = False -- Too few args\r
- | n_args == n_vals_wanted = interesting_cont\r
- | otherwise = True -- Extra args\r
- -- This rather elaborate defn for really_interesting_cont is important\r
- -- Consider an I# = INLINE (\x -> I# {x})\r
- -- The unfolding guidance deems it to have size 2, and no arguments.\r
- -- So in an application (I# y) we must take the extra arg 'y' as\r
- -- evidene of an interesting context!\r
- \r
- small_enough = (size - discount) <= opt_UF_UseThreshold\r
- discount = computeDiscount n_vals_wanted arg_discounts res_discount \r
- arg_infos really_interesting_cont\r
-\r
- \r
- in \r
-#ifdef DEBUG\r
- if opt_D_dump_inlinings then\r
- pprTrace "Considering inlining"\r
- (ppr id <+> vcat [text "black listed" <+> ppr black_listed,\r
- text "inline prag:" <+> ppr inline_prag,\r
- text "arg infos" <+> ppr arg_infos,\r
- text "interesting continuation" <+> ppr interesting_cont,\r
- text "whnf" <+> ppr whnf,\r
- text "guidance" <+> ppr guidance,\r
- text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",\r
- if yes_or_no then\r
- text "Unfolding =" <+> pprCoreExpr unf_template\r
- else empty])\r
- result\r
- else\r
-#endif\r
- result\r
- }\r
-\r
--- An argument is interesting if it has *some* structure\r
--- We are here trying to avoid unfolding a function that\r
--- is applied only to variables that have no unfolding\r
--- (i.e. they are probably lambda bound): f x y z\r
--- There is little point in inlining f here.\r
-interestingArg (Type _) = False\r
-interestingArg (App fn (Type _)) = interestingArg fn\r
-interestingArg (Var v) = hasUnfolding (getIdUnfolding v)\r
-interestingArg other = True\r
-\r
-\r
-computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int\r
-computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used\r
- -- We multiple the raw discounts (args_discount and result_discount)\r
- -- ty opt_UnfoldingKeenessFactor because the former have to do with\r
- -- *size* whereas the discounts imply that there's some extra \r
- -- *efficiency* to be gained (e.g. beta reductions, case reductions) \r
- -- by inlining.\r
-\r
- -- we also discount 1 for each argument passed, because these will\r
- -- reduce with the lambdas in the function (we count 1 for a lambda\r
- -- in size_up).\r
- = length (take n_vals_wanted arg_infos) +\r
- -- Discount of 1 for each arg supplied, because the \r
- -- result replaces the call\r
- round (opt_UF_KeenessFactor * \r
- fromInt (arg_discount + result_discount))\r
- where\r
- arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)\r
-\r
- mk_arg_discount discount is_evald | is_evald = discount\r
- | otherwise = 0\r
-\r
- -- Don't give a result discount unless there are enough args\r
- result_discount | result_used = res_discount -- Over-applied, or case scrut\r
- | otherwise = 0\r
-\end{code}\r
-\r
-\r
-%************************************************************************\r
-%* *\r
-\subsection{Black-listing}\r
-%* *\r
-%************************************************************************\r
-\r
-Inlining is controlled by the "Inline phase" number, which is set\r
-by the per-simplification-pass '-finline-phase' flag.\r
-\r
-For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)\r
-in that order. The meanings of these are determined by the @blackListed@ function\r
-here.\r
-\r
-\begin{code}\r
-blackListed :: IdSet -- Used in transformation rules\r
- -> Maybe Int -- Inline phase\r
- -> Id -> Bool -- True <=> blacklisted\r
- \r
--- The blackListed function sees whether a variable should *not* be \r
--- inlined because of the inline phase we are in. This is the sole\r
--- place that the inline phase number is looked at.\r
-\r
--- Phase 0: used for 'no inlinings please'\r
-blackListed rule_vars (Just 0)\r
- = \v -> True\r
-\r
--- Phase 1: don't inline any rule-y things or things with specialisations\r
-blackListed rule_vars (Just 1)\r
- = \v -> let v_uniq = idUnique v\r
- in v `elemVarSet` rule_vars\r
- || not (isEmptyCoreRules (getIdSpecialisation v))\r
- || v_uniq == runSTRepIdKey\r
-\r
--- Phase 2: allow build/augment to inline, and specialisations\r
-blackListed rule_vars (Just 2)\r
- = \v -> let v_uniq = idUnique v\r
- in (v `elemVarSet` rule_vars && not (v_uniq == buildIdKey || \r
- v_uniq == augmentIdKey))\r
- || v_uniq == runSTRepIdKey\r
-\r
--- Otherwise just go for it\r
-blackListed rule_vars phase\r
- = \v -> False\r
-\end{code}\r
-\r
-\r
-SLPJ 95/04: Why @runST@ must be inlined very late:\r
-\begin{verbatim}\r
-f x =\r
- runST ( \ s -> let\r
- (a, s') = newArray# 100 [] s\r
- (_, s'') = fill_in_array_or_something a x s'\r
- in\r
- freezeArray# a s'' )\r
-\end{verbatim}\r
-If we inline @runST@, we'll get:\r
-\begin{verbatim}\r
-f x = let\r
- (a, s') = newArray# 100 [] realWorld#{-NB-}\r
- (_, s'') = fill_in_array_or_something a x s'\r
- in\r
- freezeArray# a s''\r
-\end{verbatim}\r
-And now the @newArray#@ binding can be floated to become a CAF, which\r
-is totally and utterly wrong:\r
-\begin{verbatim}\r
-f = let\r
- (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!\r
- in\r
- \ x ->\r
- let (_, s'') = fill_in_array_or_something a x s' in\r
- freezeArray# a s''\r
-\end{verbatim}\r
-All calls to @f@ will share a {\em single} array! \r
-\r
-Yet we do want to inline runST sometime, so we can avoid\r
-needless code. Solution: black list it until the last moment.\r
-\r
+%
+% (c) The AQUA Project, Glasgow University, 1994-1998
+%
+\section[CoreUnfold]{Core-syntax unfoldings}
+
+Unfoldings (which can travel across module boundaries) are in Core
+syntax (namely @CoreExpr@s).
+
+The type @Unfolding@ sits ``above'' simply-Core-expressions
+unfoldings, capturing ``higher-level'' things we know about a binding,
+usually things that the simplifier found out (e.g., ``it's a
+literal''). In the corner of a @CoreUnfolding@ unfolding, you will
+find, unsurprisingly, a Core expression.
+
+\begin{code}
+module CoreUnfold (
+ Unfolding, UnfoldingGuidance, -- Abstract types
+
+ noUnfolding, mkTopUnfolding, mkUnfolding, mkCompulsoryUnfolding, seqUnfolding,
+ mkOtherCon, otherCons,
+ unfoldingTemplate, maybeUnfoldingTemplate,
+ isEvaldUnfolding, isValueUnfolding, isCheapUnfolding, isCompulsoryUnfolding,
+ hasUnfolding, hasSomeUnfolding, neverUnfold,
+
+ couldBeSmallEnoughToInline,
+ certainlyWillInline,
+ okToUnfoldInHiFile,
+
+ callSiteInline, blackListed
+ ) where
+
+#include "HsVersions.h"
+
+import CmdLineOpts ( opt_UF_CreationThreshold,
+ opt_UF_UseThreshold,
+ opt_UF_FunAppDiscount,
+ opt_UF_KeenessFactor,
+ opt_UF_DearOp, opt_UnfoldCasms,
+ DynFlags, DynFlag(..), dopt
+ )
+import CoreSyn
+import PprCore ( pprCoreExpr )
+import OccurAnal ( occurAnalyseGlobalExpr )
+import CoreUtils ( exprIsValue, exprIsCheap, exprIsTrivial )
+import Id ( Id, idType, idFlavour, isId,
+ idSpecialisation, idInlinePragma, idUnfolding,
+ isPrimOpId_maybe
+ )
+import VarSet
+import Literal ( isLitLitLit, litIsDupable )
+import PrimOp ( PrimOp(..), primOpIsDupable, primOpOutOfLine, ccallIsCasm )
+import IdInfo ( InlinePragInfo(..), OccInfo(..), IdFlavour(..),
+ isNeverInlinePrag
+ )
+import Type ( isUnLiftedType )
+import PrelNames ( hasKey, buildIdKey, augmentIdKey )
+import Bag
+import FastTypes
+import Outputable
+
+#if __GLASGOW_HASKELL__ >= 404
+import GlaExts ( fromInt )
+#endif
+\end{code}
+
+
+%************************************************************************
+%* *
+\subsection{Making unfoldings}
+%* *
+%************************************************************************
+
+\begin{code}
+mkTopUnfolding expr = mkUnfolding True {- Top level -} expr
+
+mkUnfolding top_lvl expr
+ = CoreUnfolding (occurAnalyseGlobalExpr expr)
+ top_lvl
+ (exprIsValue expr)
+ -- Already evaluated
+
+ (exprIsCheap expr)
+ -- OK to inline inside a lambda
+
+ (calcUnfoldingGuidance opt_UF_CreationThreshold expr)
+ -- Sometimes during simplification, there's a large let-bound thing
+ -- which has been substituted, and so is now dead; so 'expr' contains
+ -- two copies of the thing while the occurrence-analysed expression doesn't
+ -- Nevertheless, we don't occ-analyse before computing the size because the
+ -- size computation bales out after a while, whereas occurrence analysis does not.
+ --
+ -- This can occasionally mean that the guidance is very pessimistic;
+ -- it gets fixed up next round
+
+mkCompulsoryUnfolding expr -- Used for things that absolutely must be unfolded
+ = CompulsoryUnfolding (occurAnalyseGlobalExpr expr)
+\end{code}
+
+
+%************************************************************************
+%* *
+\subsection{The UnfoldingGuidance type}
+%* *
+%************************************************************************
+
+\begin{code}
+instance Outputable UnfoldingGuidance where
+ ppr UnfoldNever = ptext SLIT("NEVER")
+ ppr (UnfoldIfGoodArgs v cs size discount)
+ = hsep [ ptext SLIT("IF_ARGS"), int v,
+ brackets (hsep (map int cs)),
+ int size,
+ int discount ]
+\end{code}
+
+
+\begin{code}
+calcUnfoldingGuidance
+ :: Int -- bomb out if size gets bigger than this
+ -> CoreExpr -- expression to look at
+ -> UnfoldingGuidance
+calcUnfoldingGuidance bOMB_OUT_SIZE expr
+ = case collect_val_bndrs expr of { (inline, val_binders, body) ->
+ let
+ n_val_binders = length val_binders
+
+ max_inline_size = n_val_binders+2
+ -- The idea is that if there is an INLINE pragma (inline is True)
+ -- and there's a big body, we give a size of n_val_binders+2. This
+ -- This is just enough to fail the no-size-increase test in callSiteInline,
+ -- so that INLINE things don't get inlined into entirely boring contexts,
+ -- but no more.
+
+ in
+ case (sizeExpr bOMB_OUT_SIZE val_binders body) of
+
+ TooBig
+ | not inline -> UnfoldNever
+ -- A big function with an INLINE pragma must
+ -- have an UnfoldIfGoodArgs guidance
+ | inline -> UnfoldIfGoodArgs n_val_binders
+ (map (const 0) val_binders)
+ max_inline_size 0
+
+ SizeIs size cased_args scrut_discount
+ -> UnfoldIfGoodArgs
+ n_val_binders
+ (map discount_for val_binders)
+ final_size
+ (iBox scrut_discount)
+ where
+ boxed_size = iBox size
+
+ final_size | inline = boxed_size `min` max_inline_size
+ | otherwise = boxed_size
+
+ -- Sometimes an INLINE thing is smaller than n_val_binders+2.
+ -- A particular case in point is a constructor, which has size 1.
+ -- We want to inline this regardless, hence the `min`
+
+ discount_for b = foldlBag (\acc (b',n) -> if b==b' then acc+n else acc)
+ 0 cased_args
+ }
+ where
+ collect_val_bndrs e = go False [] e
+ -- We need to be a bit careful about how we collect the
+ -- value binders. In ptic, if we see
+ -- __inline_me (\x y -> e)
+ -- We want to say "2 value binders". Why? So that
+ -- we take account of information given for the arguments
+
+ go inline rev_vbs (Note InlineMe e) = go True rev_vbs e
+ go inline rev_vbs (Lam b e) | isId b = go inline (b:rev_vbs) e
+ | otherwise = go inline rev_vbs e
+ go inline rev_vbs e = (inline, reverse rev_vbs, e)
+\end{code}
+
+\begin{code}
+sizeExpr :: Int -- Bomb out if it gets bigger than this
+ -> [Id] -- Arguments; we're interested in which of these
+ -- get case'd
+ -> CoreExpr
+ -> ExprSize
+
+sizeExpr bOMB_OUT_SIZE top_args expr
+ = size_up expr
+ where
+ size_up (Type t) = sizeZero -- Types cost nothing
+ size_up (Var v) = sizeOne
+
+ size_up (Note _ body) = size_up body -- Notes cost nothing
+
+ size_up (App fun (Type t)) = size_up fun
+ size_up (App fun arg) = size_up_app fun [arg]
+
+ size_up (Lit lit) | litIsDupable lit = sizeOne
+ | otherwise = sizeN opt_UF_DearOp -- For lack of anything better
+
+ size_up (Lam b e) | isId b = lamScrutDiscount (size_up e `addSizeN` 1)
+ | otherwise = size_up e
+
+ size_up (Let (NonRec binder rhs) body)
+ = nukeScrutDiscount (size_up rhs) `addSize`
+ size_up body `addSizeN`
+ (if isUnLiftedType (idType binder) then 0 else 1)
+ -- For the allocation
+ -- If the binder has an unlifted type there is no allocation
+
+ size_up (Let (Rec pairs) body)
+ = nukeScrutDiscount rhs_size `addSize`
+ size_up body `addSizeN`
+ length pairs -- For the allocation
+ where
+ rhs_size = foldr (addSize . size_up . snd) sizeZero pairs
+
+ size_up (Case (Var v) _ alts)
+ | v `elem` top_args -- We are scrutinising an argument variable
+ =
+{- I'm nuking this special case; BUT see the comment with case alternatives.
+
+ (a) It's too eager. We don't want to inline a wrapper into a
+ context with no benefit.
+ E.g. \ x. f (x+x) no point in inlining (+) here!
+
+ (b) It's ineffective. Once g's wrapper is inlined, its case-expressions
+ aren't scrutinising arguments any more
+
+ case alts of
+
+ [alt] -> size_up_alt alt `addSize` SizeIs 0# (unitBag (v, 1)) 0#
+ -- We want to make wrapper-style evaluation look cheap, so that
+ -- when we inline a wrapper it doesn't make call site (much) bigger
+ -- Otherwise we get nasty phase ordering stuff:
+ -- f x = g x x
+ -- h y = ...(f e)...
+ -- If we inline g's wrapper, f looks big, and doesn't get inlined
+ -- into h; if we inline f first, while it looks small, then g's
+ -- wrapper will get inlined later anyway. To avoid this nasty
+ -- ordering difference, we make (case a of (x,y) -> ...),
+ -- *where a is one of the arguments* look free.
+
+ other ->
+-}
+ alts_size (foldr addSize sizeOne alt_sizes) -- The 1 is for the scrutinee
+ (foldr1 maxSize alt_sizes)
+
+ -- Good to inline if an arg is scrutinised, because
+ -- that may eliminate allocation in the caller
+ -- And it eliminates the case itself
+
+ where
+ alt_sizes = map size_up_alt alts
+
+ -- alts_size tries to compute a good discount for
+ -- the case when we are scrutinising an argument variable
+ alts_size (SizeIs tot tot_disc tot_scrut) -- Size of all alternatives
+ (SizeIs max max_disc max_scrut) -- Size of biggest alternative
+ = SizeIs tot (unitBag (v, iBox (_ILIT 1 +# tot -# max)) `unionBags` max_disc) max_scrut
+ -- If the variable is known, we produce a discount that
+ -- will take us back to 'max', the size of rh largest alternative
+ -- The 1+ is a little discount for reduced allocation in the caller
+ alts_size tot_size _ = tot_size
+
+
+ size_up (Case e _ alts) = nukeScrutDiscount (size_up e) `addSize`
+ foldr (addSize . size_up_alt) sizeZero alts
+ -- We don't charge for the case itself
+ -- It's a strict thing, and the price of the call
+ -- is paid by scrut. Also consider
+ -- case f x of DEFAULT -> e
+ -- This is just ';'! Don't charge for it.
+
+ ------------
+ size_up_app (App fun arg) args
+ | isTypeArg arg = size_up_app fun args
+ | otherwise = size_up_app fun (arg:args)
+ size_up_app fun args = foldr (addSize . nukeScrutDiscount . size_up)
+ (size_up_fun fun args)
+ args
+
+ -- A function application with at least one value argument
+ -- so if the function is an argument give it an arg-discount
+ --
+ -- Also behave specially if the function is a build
+ --
+ -- Also if the function is a constant Id (constr or primop)
+ -- compute discounts specially
+ size_up_fun (Var fun) args
+ | fun `hasKey` buildIdKey = buildSize
+ | fun `hasKey` augmentIdKey = augmentSize
+ | otherwise
+ = case idFlavour fun of
+ DataConId dc -> conSizeN (valArgCount args)
+
+ PrimOpId op -> primOpSize op (valArgCount args)
+ -- foldr addSize (primOpSize op) (map arg_discount args)
+ -- At one time I tried giving an arg-discount if a primop
+ -- is applied to one of the function's arguments, but it's
+ -- not good. At the moment, any unlifted-type arg gets a
+ -- 'True' for 'yes I'm evald', so we collect the discount even
+ -- if we know nothing about it. And just having it in a primop
+ -- doesn't help at all if we don't know something more.
+
+ other -> fun_discount fun `addSizeN`
+ (1 + length (filter (not . exprIsTrivial) args))
+ -- The 1+ is for the function itself
+ -- Add 1 for each non-trivial arg;
+ -- the allocation cost, as in let(rec)
+ -- Slight hack here: for constructors the args are almost always
+ -- trivial; and for primops they are almost always prim typed
+ -- We should really only count for non-prim-typed args in the
+ -- general case, but that seems too much like hard work
+
+ size_up_fun other args = size_up other
+
+ ------------
+ size_up_alt (con, bndrs, rhs) = size_up rhs
+ -- Don't charge for args, so that wrappers look cheap
+ -- (See comments about wrappers with Case)
+
+ ------------
+ -- We want to record if we're case'ing, or applying, an argument
+ fun_discount v | v `elem` top_args = SizeIs 0# (unitBag (v, opt_UF_FunAppDiscount)) 0#
+ fun_discount other = sizeZero
+
+ ------------
+ -- These addSize things have to be here because
+ -- I don't want to give them bOMB_OUT_SIZE as an argument
+
+ addSizeN TooBig _ = TooBig
+ addSizeN (SizeIs n xs d) m
+ | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig
+ | otherwise = SizeIs n_tot xs d
+ where
+ n_tot = n +# iUnbox m
+
+ addSize TooBig _ = TooBig
+ addSize _ TooBig = TooBig
+ addSize (SizeIs n1 xs d1) (SizeIs n2 ys d2)
+ | n_tot ># (iUnbox bOMB_OUT_SIZE) = TooBig
+ | otherwise = SizeIs n_tot xys d_tot
+ where
+ n_tot = n1 +# n2
+ d_tot = d1 +# d2
+ xys = xs `unionBags` ys
+\end{code}
+
+Code for manipulating sizes
+
+\begin{code}
+
+data ExprSize = TooBig
+ | SizeIs FastInt -- Size found
+ (Bag (Id,Int)) -- Arguments cased herein, and discount for each such
+ FastInt -- Size to subtract if result is scrutinised
+ -- by a case expression
+
+
+maxSize TooBig _ = TooBig
+maxSize _ TooBig = TooBig
+maxSize s1@(SizeIs n1 _ _) s2@(SizeIs n2 _ _) | n1 ># n2 = s1
+ | otherwise = s2
+
+sizeZero = SizeIs (_ILIT 0) emptyBag (_ILIT 0)
+sizeOne = SizeIs (_ILIT 1) emptyBag (_ILIT 0)
+sizeTwo = SizeIs (_ILIT 2) emptyBag (_ILIT 0)
+sizeN n = SizeIs (iUnbox n) emptyBag (_ILIT 0)
+conSizeN n = SizeIs (_ILIT 1) emptyBag (iUnbox n +# _ILIT 1)
+ -- Treat constructors as size 1; we are keen to expose them
+ -- (and we charge separately for their args). We can't treat
+ -- them as size zero, else we find that (iBox x) has size 1,
+ -- which is the same as a lone variable; and hence 'v' will
+ -- always be replaced by (iBox x), where v is bound to iBox x.
+
+primOpSize op n_args
+ | not (primOpIsDupable op) = sizeN opt_UF_DearOp
+ | not (primOpOutOfLine op) = sizeN (1 - n_args)
+ -- Be very keen to inline simple primops.
+ -- We give a discount of 1 for each arg so that (op# x y z) costs 1.
+ -- I found occasions where we had
+ -- f x y z = case op# x y z of { s -> (# s, () #) }
+ -- and f wasn't getting inlined
+ | otherwise = sizeOne
+
+buildSize = SizeIs (-2#) emptyBag 4#
+ -- We really want to inline applications of build
+ -- build t (\cn -> e) should cost only the cost of e (because build will be inlined later)
+ -- Indeed, we should add a result_discount becuause build is
+ -- very like a constructor. We don't bother to check that the
+ -- build is saturated (it usually is). The "-2" discounts for the \c n,
+ -- The "4" is rather arbitrary.
+
+augmentSize = SizeIs (-2#) emptyBag 4#
+ -- Ditto (augment t (\cn -> e) ys) should cost only the cost of
+ -- e plus ys. The -2 accounts for the \cn
+
+nukeScrutDiscount (SizeIs n vs d) = SizeIs n vs 0#
+nukeScrutDiscount TooBig = TooBig
+
+-- When we return a lambda, give a discount if it's used (applied)
+lamScrutDiscount (SizeIs n vs d) = case opt_UF_FunAppDiscount of { d -> SizeIs n vs (iUnbox d) }
+lamScrutDiscount TooBig = TooBig
+\end{code}
+
+
+%************************************************************************
+%* *
+\subsection[considerUnfolding]{Given all the info, do (not) do the unfolding}
+%* *
+%************************************************************************
+
+We have very limited information about an unfolding expression: (1)~so
+many type arguments and so many value arguments expected---for our
+purposes here, we assume we've got those. (2)~A ``size'' or ``cost,''
+a single integer. (3)~An ``argument info'' vector. For this, what we
+have at the moment is a Boolean per argument position that says, ``I
+will look with great favour on an explicit constructor in this
+position.'' (4)~The ``discount'' to subtract if the expression
+is being scrutinised.
+
+Assuming we have enough type- and value arguments (if not, we give up
+immediately), then we see if the ``discounted size'' is below some
+(semi-arbitrary) threshold. It works like this: for every argument
+position where we're looking for a constructor AND WE HAVE ONE in our
+hands, we get a (again, semi-arbitrary) discount [proportion to the
+number of constructors in the type being scrutinized].
+
+If we're in the context of a scrutinee ( \tr{(case <expr > of A .. -> ...;.. )})
+and the expression in question will evaluate to a constructor, we use
+the computed discount size *for the result only* rather than
+computing the argument discounts. Since we know the result of
+the expression is going to be taken apart, discounting its size
+is more accurate (see @sizeExpr@ above for how this discount size
+is computed).
+
+We use this one to avoid exporting inlinings that we ``couldn't possibly
+use'' on the other side. Can be overridden w/ flaggery.
+Just the same as smallEnoughToInline, except that it has no actual arguments.
+
+\begin{code}
+couldBeSmallEnoughToInline :: Int -> CoreExpr -> Bool
+couldBeSmallEnoughToInline threshold rhs = case calcUnfoldingGuidance threshold rhs of
+ UnfoldNever -> False
+ other -> True
+
+certainlyWillInline :: Id -> Bool
+ -- Sees if the Id is pretty certain to inline
+certainlyWillInline v
+ = case idUnfolding v of
+
+ CoreUnfolding _ _ is_value _ g@(UnfoldIfGoodArgs n_vals _ size _)
+ -> is_value
+ && size - (n_vals +1) <= opt_UF_UseThreshold
+
+ other -> False
+\end{code}
+
+@okToUnfoldInHifile@ is used when emitting unfolding info into an interface
+file to determine whether an unfolding candidate really should be unfolded.
+The predicate is needed to prevent @_casm_@s (+ lit-lits) from being emitted
+into interface files.
+
+The reason for inlining expressions containing _casm_s into interface files
+is that these fragments of C are likely to mention functions/#defines that
+will be out-of-scope when inlined into another module. This is not an
+unfixable problem for the user (just need to -#include the approp. header
+file), but turning it off seems to the simplest thing to do.
+
+\begin{code}
+okToUnfoldInHiFile :: CoreExpr -> Bool
+okToUnfoldInHiFile e = opt_UnfoldCasms || go e
+ where
+ -- Race over an expression looking for CCalls..
+ go (Var v) = case isPrimOpId_maybe v of
+ Just op -> okToUnfoldPrimOp op
+ Nothing -> True
+ go (Lit lit) = not (isLitLitLit lit)
+ go (App fun arg) = go fun && go arg
+ go (Lam _ body) = go body
+ go (Let binds body) = and (map go (body :rhssOfBind binds))
+ go (Case scrut bndr alts) = and (map go (scrut:rhssOfAlts alts)) &&
+ not (any isLitLitLit [ lit | (LitAlt lit, _, _) <- alts ])
+ go (Note _ body) = go body
+ go (Type _) = True
+
+ -- ok to unfold a PrimOp as long as it's not a _casm_
+ okToUnfoldPrimOp (CCallOp ccall) = not (ccallIsCasm ccall)
+ okToUnfoldPrimOp _ = True
+\end{code}
+
+
+%************************************************************************
+%* *
+\subsection{callSiteInline}
+%* *
+%************************************************************************
+
+This is the key function. It decides whether to inline a variable at a call site
+
+callSiteInline is used at call sites, so it is a bit more generous.
+It's a very important function that embodies lots of heuristics.
+A non-WHNF can be inlined if it doesn't occur inside a lambda,
+and occurs exactly once or
+ occurs once in each branch of a case and is small
+
+If the thing is in WHNF, there's no danger of duplicating work,
+so we can inline if it occurs once, or is small
+
+NOTE: we don't want to inline top-level functions that always diverge.
+It just makes the code bigger. Tt turns out that the convenient way to prevent
+them inlining is to give them a NOINLINE pragma, which we do in
+StrictAnal.addStrictnessInfoToTopId
+
+\begin{code}
+callSiteInline :: DynFlags
+ -> Bool -- True <=> the Id is black listed
+ -> Bool -- 'inline' note at call site
+ -> OccInfo
+ -> Id -- The Id
+ -> [Bool] -- One for each value arg; True if it is interesting
+ -> Bool -- True <=> continuation is interesting
+ -> Maybe CoreExpr -- Unfolding, if any
+
+
+callSiteInline dflags black_listed inline_call occ id arg_infos interesting_cont
+ = case idUnfolding id of {
+ NoUnfolding -> Nothing ;
+ OtherCon cs -> Nothing ;
+ CompulsoryUnfolding unf_template | black_listed -> Nothing
+ | otherwise -> Just unf_template ;
+ -- Constructors have compulsory unfoldings, but
+ -- may have rules, in which case they are
+ -- black listed till later
+ CoreUnfolding unf_template is_top is_value is_cheap guidance ->
+
+ let
+ result | yes_or_no = Just unf_template
+ | otherwise = Nothing
+
+ n_val_args = length arg_infos
+
+ yes_or_no
+ | black_listed = False
+ | otherwise = case occ of
+ IAmDead -> pprTrace "callSiteInline: dead" (ppr id) False
+ IAmALoopBreaker -> False
+ OneOcc in_lam one_br -> (not in_lam || is_cheap) && consider_safe in_lam True one_br
+ NoOccInfo -> is_cheap && consider_safe True False False
+
+ consider_safe in_lam once once_in_one_branch
+ -- consider_safe decides whether it's a good idea to inline something,
+ -- given that there's no work-duplication issue (the caller checks that).
+ -- once_in_one_branch = True means there's a unique textual occurrence
+ | inline_call = True
+
+ | once_in_one_branch
+ -- Be very keen to inline something if this is its unique occurrence:
+ --
+ -- a) Inlining gives a good chance of eliminating the original
+ -- binding (and hence the allocation) for the thing.
+ -- (Provided it's not a top level binding, in which case the
+ -- allocation costs nothing.)
+ --
+ -- b) Inlining a function that is called only once exposes the
+ -- body function to the call site.
+ --
+ -- The only time we hold back is when substituting inside a lambda;
+ -- then if the context is totally uninteresting (not applied, not scrutinised)
+ -- there is no point in substituting because it might just increase allocation,
+ -- by allocating the function itself many times
+ --
+ -- Note: there used to be a '&& not top_level' in the guard above,
+ -- but that stopped us inlining top-level functions used only once,
+ -- which is stupid
+ = not in_lam || not (null arg_infos) || interesting_cont
+
+ | otherwise
+ = case guidance of
+ UnfoldNever -> False ;
+ UnfoldIfGoodArgs n_vals_wanted arg_discounts size res_discount
+
+ | enough_args && size <= (n_vals_wanted + 1)
+ -- No size increase
+ -- Size of call is n_vals_wanted (+1 for the function)
+ -> True
+
+ | otherwise
+ -> some_benefit && small_enough
+
+ where
+ some_benefit = or arg_infos || really_interesting_cont ||
+ (not is_top && (once || (n_vals_wanted > 0 && enough_args)))
+ -- If it occurs more than once, there must be something interesting
+ -- about some argument, or the result context, to make it worth inlining
+ --
+ -- If a function has a nested defn we also record some-benefit,
+ -- on the grounds that we are often able to eliminate the binding,
+ -- and hence the allocation, for the function altogether; this is good
+ -- for join points. But this only makes sense for *functions*;
+ -- inlining a constructor doesn't help allocation unless the result is
+ -- scrutinised. UNLESS the constructor occurs just once, albeit possibly
+ -- in multiple case branches. Then inlining it doesn't increase allocation,
+ -- but it does increase the chance that the constructor won't be allocated at all
+ -- in the branches that don't use it.
+
+ enough_args = n_val_args >= n_vals_wanted
+ really_interesting_cont | n_val_args < n_vals_wanted = False -- Too few args
+ | n_val_args == n_vals_wanted = interesting_cont
+ | otherwise = True -- Extra args
+ -- really_interesting_cont tells if the result of the
+ -- call is in an interesting context.
+
+ small_enough = (size - discount) <= opt_UF_UseThreshold
+ discount = computeDiscount n_vals_wanted arg_discounts res_discount
+ arg_infos really_interesting_cont
+
+ in
+#ifdef DEBUG
+ if dopt Opt_D_dump_inlinings dflags then
+ pprTrace "Considering inlining"
+ (ppr id <+> vcat [text "black listed:" <+> ppr black_listed,
+ text "occ info:" <+> ppr occ,
+ text "arg infos" <+> ppr arg_infos,
+ text "interesting continuation" <+> ppr interesting_cont,
+ text "is value:" <+> ppr is_value,
+ text "is cheap:" <+> ppr is_cheap,
+ text "guidance" <+> ppr guidance,
+ text "ANSWER =" <+> if yes_or_no then text "YES" else text "NO",
+ if yes_or_no then
+ text "Unfolding =" <+> pprCoreExpr unf_template
+ else empty])
+ result
+ else
+#endif
+ result
+ }
+
+computeDiscount :: Int -> [Int] -> Int -> [Bool] -> Bool -> Int
+computeDiscount n_vals_wanted arg_discounts res_discount arg_infos result_used
+ -- We multiple the raw discounts (args_discount and result_discount)
+ -- ty opt_UnfoldingKeenessFactor because the former have to do with
+ -- *size* whereas the discounts imply that there's some extra
+ -- *efficiency* to be gained (e.g. beta reductions, case reductions)
+ -- by inlining.
+
+ -- we also discount 1 for each argument passed, because these will
+ -- reduce with the lambdas in the function (we count 1 for a lambda
+ -- in size_up).
+ = 1 + -- Discount of 1 because the result replaces the call
+ -- so we count 1 for the function itself
+ length (take n_vals_wanted arg_infos) +
+ -- Discount of 1 for each arg supplied, because the
+ -- result replaces the call
+ round (opt_UF_KeenessFactor *
+ fromInt (arg_discount + result_discount))
+ where
+ arg_discount = sum (zipWith mk_arg_discount arg_discounts arg_infos)
+
+ mk_arg_discount discount is_evald | is_evald = discount
+ | otherwise = 0
+
+ -- Don't give a result discount unless there are enough args
+ result_discount | result_used = res_discount -- Over-applied, or case scrut
+ | otherwise = 0
+\end{code}
+
+
+%************************************************************************
+%* *
+\subsection{Black-listing}
+%* *
+%************************************************************************
+
+Inlining is controlled by the "Inline phase" number, which is set
+by the per-simplification-pass '-finline-phase' flag.
+
+For optimisation we use phase 1,2 and nothing (i.e. no -finline-phase flag)
+in that order. The meanings of these are determined by the @blackListed@ function
+here.
+
+The final simplification doesn't have a phase number.
+
+Pragmas
+~~~~~~~
+ Pragma Black list if
+
+(least black listing, most inlining)
+ INLINE n foo phase is Just p *and* p<n *and* foo appears on LHS of rule
+ INLINE foo phase is Just p *and* foo appears on LHS of rule
+ NOINLINE n foo phase is Just p *and* (p<n *or* foo appears on LHS of rule)
+ NOINLINE foo always
+(most black listing, least inlining)
+
+\begin{code}
+blackListed :: IdSet -- Used in transformation rules
+ -> Maybe Int -- Inline phase
+ -> Id -> Bool -- True <=> blacklisted
+
+-- The blackListed function sees whether a variable should *not* be
+-- inlined because of the inline phase we are in. This is the sole
+-- place that the inline phase number is looked at.
+
+blackListed rule_vars Nothing -- Last phase
+ = \v -> isNeverInlinePrag (idInlinePragma v)
+
+blackListed rule_vars (Just phase)
+ = \v -> normal_case rule_vars phase v
+
+normal_case rule_vars phase v
+ = case idInlinePragma v of
+ NoInlinePragInfo -> has_rules
+
+ IMustNotBeINLINEd from_INLINE Nothing
+ | from_INLINE -> has_rules -- Black list until final phase
+ | otherwise -> True -- Always blacklisted
+
+ IMustNotBeINLINEd from_INLINE (Just threshold)
+ | from_INLINE -> (phase < threshold && has_rules)
+ | otherwise -> (phase < threshold || has_rules)
+ where
+ has_rules = v `elemVarSet` rule_vars
+ || not (isEmptyCoreRules (idSpecialisation v))
+\end{code}
+
+
+SLPJ 95/04: Why @runST@ must be inlined very late:
+\begin{verbatim}
+f x =
+ runST ( \ s -> let
+ (a, s') = newArray# 100 [] s
+ (_, s'') = fill_in_array_or_something a x s'
+ in
+ freezeArray# a s'' )
+\end{verbatim}
+If we inline @runST@, we'll get:
+\begin{verbatim}
+f x = let
+ (a, s') = newArray# 100 [] realWorld#{-NB-}
+ (_, s'') = fill_in_array_or_something a x s'
+ in
+ freezeArray# a s''
+\end{verbatim}
+And now the @newArray#@ binding can be floated to become a CAF, which
+is totally and utterly wrong:
+\begin{verbatim}
+f = let
+ (a, s') = newArray# 100 [] realWorld#{-NB-} -- YIKES!!!
+ in
+ \ x ->
+ let (_, s'') = fill_in_array_or_something a x s' in
+ freezeArray# a s''
+\end{verbatim}
+All calls to @f@ will share a {\em single} array!
+
+Yet we do want to inline runST sometime, so we can avoid
+needless code. Solution: black list it until the last moment.
+
projects / ghc-hetmet.git / blobdiff