CmmNode.hs 29.2 KB
Newer Older
1 2
{-# LANGUAGE CPP #-}
{-# LANGUAGE ExplicitForAll #-}
3
{-# LANGUAGE FlexibleContexts #-}
4
{-# LANGUAGE FlexibleInstances #-}
5
{-# LANGUAGE GADTs #-}
6
{-# LANGUAGE MultiParamTypeClasses #-}
Gabor Greif's avatar
Gabor Greif committed
7
{-# LANGUAGE StandaloneDeriving #-}
8
{-# LANGUAGE UndecidableInstances #-}
Simon Marlow's avatar
Simon Marlow committed
9

10 11
-- CmmNode type for representation using Hoopl graphs.

12
module CmmNode (
Peter Wortmann's avatar
Peter Wortmann committed
13
     CmmNode(..), CmmFormal, CmmActual, CmmTickish,
14 15
     UpdFrameOffset, Convention(..),
     ForeignConvention(..), ForeignTarget(..), foreignTargetHints,
16
     CmmReturnInfo(..),
17
     mapExp, mapExpDeep, wrapRecExp, foldExp, foldExpDeep, wrapRecExpf,
Peter Wortmann's avatar
Peter Wortmann committed
18 19 20 21
     mapExpM, mapExpDeepM, wrapRecExpM, mapSuccessors,

     -- * Tick scopes
     CmmTickScope(..), isTickSubScope, combineTickScopes,
22
  ) where
23

24
import CodeGen.Platform
25
import CmmExpr
26
import DynFlags
27 28
import FastString
import ForeignCall
Peter Wortmann's avatar
Peter Wortmann committed
29
import Outputable
30
import SMRep
Peter Wortmann's avatar
Peter Wortmann committed
31
import CoreSyn (Tickish)
Peter Wortmann's avatar
Peter Wortmann committed
32
import qualified Unique as U
33 34 35

import Compiler.Hoopl
import Data.Maybe
Peter Wortmann's avatar
Peter Wortmann committed
36
import Data.List (tails,sort)
37 38 39 40 41 42
import Prelude hiding (succ)


------------------------
-- CmmNode

Simon Marlow's avatar
Simon Marlow committed
43 44
#define ULabel {-# UNPACK #-} !Label

45
data CmmNode e x where
Peter Wortmann's avatar
Peter Wortmann committed
46
  CmmEntry :: ULabel -> CmmTickScope -> CmmNode C O
Simon Peyton Jones's avatar
Simon Peyton Jones committed
47

48
  CmmComment :: FastString -> CmmNode O O
Simon Peyton Jones's avatar
Simon Peyton Jones committed
49

Peter Wortmann's avatar
Peter Wortmann committed
50 51
    -- Tick annotation, covering Cmm code in our tick scope. We only
    -- expect non-code @Tickish@ at this point (e.g. @SourceNote@).
Peter Wortmann's avatar
Peter Wortmann committed
52
    -- See Note [CmmTick scoping details]
Peter Wortmann's avatar
Peter Wortmann committed
53 54
  CmmTick :: !CmmTickish -> CmmNode O O

Simon Marlow's avatar
Simon Marlow committed
55
  CmmAssign :: !CmmReg -> !CmmExpr -> CmmNode O O
Simon Marlow's avatar
Simon Marlow committed
56
    -- Assign to register
Simon Peyton Jones's avatar
Simon Peyton Jones committed
57

Simon Marlow's avatar
Simon Marlow committed
58
  CmmStore :: !CmmExpr -> !CmmExpr -> CmmNode O O
Simon Marlow's avatar
Simon Marlow committed
59 60
    -- Assign to memory location.  Size is
    -- given by cmmExprType of the rhs.
Simon Peyton Jones's avatar
Simon Peyton Jones committed
61

62 63 64 65 66 67 68
  CmmUnsafeForeignCall ::       -- An unsafe foreign call;
                                -- see Note [Foreign calls]
                                -- Like a "fat machine instruction"; can occur
                                -- in the middle of a block
      ForeignTarget ->          -- call target
      [CmmFormal] ->            -- zero or more results
      [CmmActual] ->            -- zero or more arguments
69
      CmmNode O O
70
      -- Semantics: clobbers any GlobalRegs for which callerSaves r == True
71
      -- See Note [Unsafe foreign calls clobber caller-save registers]
72
      --
Simon Marlow's avatar
Simon Marlow committed
73 74 75
      -- Invariant: the arguments and the ForeignTarget must not
      -- mention any registers for which CodeGen.Platform.callerSaves
      -- is True.  See Note [Register Parameter Passing].
Simon Peyton Jones's avatar
Simon Peyton Jones committed
76

Simon Marlow's avatar
Simon Marlow committed
77 78
  CmmBranch :: ULabel -> CmmNode O C
                                   -- Goto another block in the same procedure
Simon Peyton Jones's avatar
Simon Peyton Jones committed
79

80 81
  CmmCondBranch :: {                 -- conditional branch
      cml_pred :: CmmExpr,
Simon Marlow's avatar
Simon Marlow committed
82
      cml_true, cml_false :: ULabel
83
  } -> CmmNode O C
Simon Peyton Jones's avatar
Simon Peyton Jones committed
84

85 86 87 88 89
  CmmSwitch :: CmmExpr -> [Maybe Label] -> CmmNode O C -- Table branch
      -- The scrutinee is zero-based;
      --      zero -> first block
      --      one  -> second block etc
      -- Undefined outside range, and when there's a Nothing
Simon Peyton Jones's avatar
Simon Peyton Jones committed
90 91

  CmmCall :: {                -- A native call or tail call
92 93 94 95
      cml_target :: CmmExpr,  -- never a CmmPrim to a CallishMachOp!

      cml_cont :: Maybe Label,
          -- Label of continuation (Nothing for return or tail call)
96 97 98 99 100
          --
          -- Note [Continuation BlockId]: these BlockIds are called
          -- Continuation BlockIds, and are the only BlockIds that can
          -- occur in CmmExprs, namely as (CmmLit (CmmBlock b)) or
          -- (CmmStackSlot (Young b) _).
101

102 103 104 105 106 107 108 109
      cml_args_regs :: [GlobalReg],
          -- The argument GlobalRegs (Rx, Fx, Dx, Lx) that are passed
          -- to the call.  This is essential information for the
          -- native code generator's register allocator; without
          -- knowing which GlobalRegs are live it has to assume that
          -- they are all live.  This list should only include
          -- GlobalRegs that are mapped to real machine registers on
          -- the target platform.
Simon Peyton Jones's avatar
Simon Peyton Jones committed
110

111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
      cml_args :: ByteOff,
          -- Byte offset, from the *old* end of the Area associated with
          -- the Label (if cml_cont = Nothing, then Old area), of
          -- youngest outgoing arg.  Set the stack pointer to this before
          -- transferring control.
          -- (NB: an update frame might also have been stored in the Old
          --      area, but it'll be in an older part than the args.)

      cml_ret_args :: ByteOff,
          -- For calls *only*, the byte offset for youngest returned value
          -- This is really needed at the *return* point rather than here
          -- at the call, but in practice it's convenient to record it here.

      cml_ret_off :: ByteOff
        -- For calls *only*, the byte offset of the base of the frame that
        -- must be described by the info table for the return point.
        -- The older words are an update frames, which have their own
        -- info-table and layout information

        -- From a liveness point of view, the stack words older than
        -- cml_ret_off are treated as live, even if the sequel of
        -- the call goes into a loop.
  } -> CmmNode O C
Simon Peyton Jones's avatar
Simon Peyton Jones committed
134

135
  CmmForeignCall :: {           -- A safe foreign call; see Note [Foreign calls]
136
                                -- Always the last node of a block
137
      tgt   :: ForeignTarget,   -- call target and convention
138 139
      res   :: [CmmFormal],     -- zero or more results
      args  :: [CmmActual],     -- zero or more arguments; see Note [Register parameter passing]
Simon Marlow's avatar
Simon Marlow committed
140
      succ  :: ULabel,          -- Label of continuation
141 142
      ret_args :: ByteOff,      -- same as cml_ret_args
      ret_off :: ByteOff,       -- same as cml_ret_off
143 144 145 146 147
      intrbl:: Bool             -- whether or not the call is interruptible
  } -> CmmNode O C

{- Note [Foreign calls]
~~~~~~~~~~~~~~~~~~~~~~~
Simon Peyton Jones's avatar
Simon Peyton Jones committed
148 149
A CmmUnsafeForeignCall is used for *unsafe* foreign calls;
a CmmForeignCall call is used for *safe* foreign calls.
150 151 152 153 154

Unsafe ones are mostly easy: think of them as a "fat machine
instruction".  In particular, they do *not* kill all live registers,
just the registers they return to (there was a bit of code in GHC that
conservatively assumed otherwise.)  However, see [Register parameter passing].
155

156
Safe ones are trickier.  A safe foreign call
157 158
     r = f(x)
ultimately expands to
159 160
     push "return address"      -- Never used to return to;
                                -- just points an info table
161 162
     save registers into TSO
     call suspendThread
163
     r = f(x)                   -- Make the call
164 165 166 167 168 169 170
     call resumeThread
     restore registers
     pop "return address"
We cannot "lower" a safe foreign call to this sequence of Cmms, because
after we've saved Sp all the Cmm optimiser's assumptions are broken.

Note that a safe foreign call needs an info table.
171 172 173 174 175 176

So Safe Foreign Calls must remain as last nodes until the stack is
made manifest in CmmLayoutStack, where they are lowered into the above
sequence.
-}

177 178
{- Note [Unsafe foreign calls clobber caller-save registers]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
179 180 181 182 183 184 185 186 187 188 189 190 191 192 193

A foreign call is defined to clobber any GlobalRegs that are mapped to
caller-saves machine registers (according to the prevailing C ABI).
StgCmmUtils.callerSaves tells you which GlobalRegs are caller-saves.

This is a design choice that makes it easier to generate code later.
We could instead choose to say that foreign calls do *not* clobber
caller-saves regs, but then we would have to figure out which regs
were live across the call later and insert some saves/restores.

Furthermore when we generate code we never have any GlobalRegs live
across a call, because they are always copied-in to LocalRegs and
copied-out again before making a call/jump.  So all we have to do is
avoid any code motion that would make a caller-saves GlobalReg live
across a foreign call during subsequent optimisations.
194 195
-}

196 197 198 199 200 201 202 203 204 205 206 207 208 209 210
{- Note [Register parameter passing]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
On certain architectures, some registers are utilized for parameter
passing in the C calling convention.  For example, in x86-64 Linux
convention, rdi, rsi, rdx and rcx (as well as r8 and r9) may be used for
argument passing.  These are registers R3-R6, which our generated
code may also be using; as a result, it's necessary to save these
values before doing a foreign call.  This is done during initial
code generation in callerSaveVolatileRegs in StgCmmUtils.hs.  However,
one result of doing this is that the contents of these registers
may mysteriously change if referenced inside the arguments.  This
is dangerous, so you'll need to disable inlining much in the same
way is done in cmm/CmmOpt.hs currently.  We should fix this!
-}

211 212
---------------------------------------------
-- Eq instance of CmmNode
Gabor Greif's avatar
Gabor Greif committed
213 214

deriving instance Eq (CmmNode e x)
215 216 217 218 219

----------------------------------------------
-- Hoopl instances of CmmNode

instance NonLocal CmmNode where
Peter Wortmann's avatar
Peter Wortmann committed
220
  entryLabel (CmmEntry l _) = l
221 222 223 224 225 226 227 228 229 230 231

  successors (CmmBranch l) = [l]
  successors (CmmCondBranch {cml_true=t, cml_false=f}) = [f, t] -- meets layout constraint
  successors (CmmSwitch _ ls) = catMaybes ls
  successors (CmmCall {cml_cont=l}) = maybeToList l
  successors (CmmForeignCall {succ=l}) = [l]


--------------------------------------------------
-- Various helper types

232 233 234
type CmmActual = CmmExpr
type CmmFormal = LocalReg

235 236
type UpdFrameOffset = ByteOff

237 238
-- | A convention maps a list of values (function arguments or return
-- values) to registers or stack locations.
239
data Convention
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261
  = NativeDirectCall
       -- ^ top-level Haskell functions use @NativeDirectCall@, which
       -- maps arguments to registers starting with R2, according to
       -- how many registers are available on the platform.  This
       -- convention ignores R1, because for a top-level function call
       -- the function closure is implicit, and doesn't need to be passed.
  | NativeNodeCall
       -- ^ non-top-level Haskell functions, which pass the address of
       -- the function closure in R1 (regardless of whether R1 is a
       -- real register or not), and the rest of the arguments in
       -- registers or on the stack.
  | NativeReturn
       -- ^ a native return.  The convention for returns depends on
       -- how many values are returned: for just one value returned,
       -- the appropriate register is used (R1, F1, etc.). regardless
       -- of whether it is a real register or not.  For multiple
       -- values returned, they are mapped to registers or the stack.
  | Slow
       -- ^ Slow entry points: all args pushed on the stack
  | GC
       -- ^ Entry to the garbage collector: uses the node reg!
       -- (TODO: I don't think we need this --SDM)
262 263 264 265 266 267 268
  deriving( Eq )

data ForeignConvention
  = ForeignConvention
        CCallConv               -- Which foreign-call convention
        [ForeignHint]           -- Extra info about the args
        [ForeignHint]           -- Extra info about the result
269
        CmmReturnInfo
270 271
  deriving Eq

272 273 274 275 276
data CmmReturnInfo
  = CmmMayReturn
  | CmmNeverReturns
  deriving ( Eq )

277 278 279 280 281 282 283 284
data ForeignTarget        -- The target of a foreign call
  = ForeignTarget                -- A foreign procedure
        CmmExpr                  -- Its address
        ForeignConvention        -- Its calling convention
  | PrimTarget            -- A possibly-side-effecting machine operation
        CallishMachOp            -- Which one
  deriving Eq

285 286 287 288 289 290 291 292 293 294 295
foreignTargetHints :: ForeignTarget -> ([ForeignHint], [ForeignHint])
foreignTargetHints target
  = ( res_hints ++ repeat NoHint
    , arg_hints ++ repeat NoHint )
  where
    (res_hints, arg_hints) =
       case target of
          PrimTarget op -> callishMachOpHints op
          ForeignTarget _ (ForeignConvention _ arg_hints res_hints _) ->
             (res_hints, arg_hints)

296 297 298
--------------------------------------------------
-- Instances of register and slot users / definers

299 300
instance UserOfRegs LocalReg (CmmNode e x) where
  foldRegsUsed dflags f z n = case n of
301 302 303 304 305 306 307 308 309
    CmmAssign _ expr -> fold f z expr
    CmmStore addr rval -> fold f (fold f z addr) rval
    CmmUnsafeForeignCall t _ args -> fold f (fold f z t) args
    CmmCondBranch expr _ _ -> fold f z expr
    CmmSwitch expr _ -> fold f z expr
    CmmCall {cml_target=tgt} -> fold f z tgt
    CmmForeignCall {tgt=tgt, args=args} -> fold f (fold f z tgt) args
    _ -> z
    where fold :: forall a b.
310
                       UserOfRegs LocalReg a =>
311
                       (b -> LocalReg -> b) -> b -> a -> b
312
          fold f z n = foldRegsUsed dflags f z n
313

314 315 316 317 318 319 320 321 322 323 324 325 326 327
instance UserOfRegs GlobalReg (CmmNode e x) where
  foldRegsUsed dflags f z n = case n of
    CmmAssign _ expr -> fold f z expr
    CmmStore addr rval -> fold f (fold f z addr) rval
    CmmUnsafeForeignCall t _ args -> fold f (fold f z t) args
    CmmCondBranch expr _ _ -> fold f z expr
    CmmSwitch expr _ -> fold f z expr
    CmmCall {cml_target=tgt, cml_args_regs=args} -> fold f (fold f z args) tgt
    CmmForeignCall {tgt=tgt, args=args} -> fold f (fold f z tgt) args
    _ -> z
    where fold :: forall a b.
                       UserOfRegs GlobalReg a =>
                       (b -> GlobalReg -> b) -> b -> a -> b
          fold f z n = foldRegsUsed dflags f z n
328

329 330 331 332 333 334
instance UserOfRegs r CmmExpr => UserOfRegs r ForeignTarget where
  foldRegsUsed _      _ z (PrimTarget _)      = z
  foldRegsUsed dflags f z (ForeignTarget e _) = foldRegsUsed dflags f z e

instance DefinerOfRegs LocalReg (CmmNode e x) where
  foldRegsDefd dflags f z n = case n of
335 336 337 338 339
    CmmAssign lhs _ -> fold f z lhs
    CmmUnsafeForeignCall _ fs _ -> fold f z fs
    CmmForeignCall {res=res} -> fold f z res
    _ -> z
    where fold :: forall a b.
340
                   DefinerOfRegs LocalReg a =>
341
                   (b -> LocalReg -> b) -> b -> a -> b
342 343 344 345 346 347
          fold f z n = foldRegsDefd dflags f z n

instance DefinerOfRegs GlobalReg (CmmNode e x) where
  foldRegsDefd dflags f z n = case n of
    CmmAssign lhs _ -> fold f z lhs
    CmmUnsafeForeignCall tgt _ _  -> fold f z (foreignTargetRegs tgt)
348 349 350
    CmmCall        {} -> fold f z activeRegs
    CmmForeignCall {} -> fold f z activeRegs
                      -- See Note [Safe foreign calls clobber STG registers]
351 352 353 354 355 356 357 358 359 360 361 362
    _ -> z
    where fold :: forall a b.
                   DefinerOfRegs GlobalReg a =>
                   (b -> GlobalReg -> b) -> b -> a -> b
          fold f z n = foldRegsDefd dflags f z n

          platform = targetPlatform dflags
          activeRegs = activeStgRegs platform
          activeCallerSavesRegs = filter (callerSaves platform) activeRegs

          foreignTargetRegs (ForeignTarget _ (ForeignConvention _ _ _ CmmNeverReturns)) = []
          foreignTargetRegs _ = activeCallerSavesRegs
363

364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431
-- Note [Safe foreign calls clobber STG registers]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- During stack layout phase every safe foreign call is expanded into a block
-- that contains unsafe foreign call (instead of safe foreign call) and ends
-- with a normal call (See Note [Foreign calls]). This means that we must
-- treat safe foreign call as if it was a normal call (because eventually it
-- will be). This is important if we try to run sinking pass before stack
-- layout phase. Consider this example of what might go wrong (this is cmm
-- code from stablename001 test). Here is code after common block elimination
-- (before stack layout):
--
--  c1q6:
--      _s1pf::P64 = R1;
--      _c1q8::I64 = performMajorGC;
--      I64[(young<c1q9> + 8)] = c1q9;
--      foreign call "ccall" arg hints:  []  result hints:  [] (_c1q8::I64)(...)
--                   returns to c1q9 args: ([]) ress: ([])ret_args: 8ret_off: 8;
--  c1q9:
--      I64[(young<c1qb> + 8)] = c1qb;
--      R1 = _s1pc::P64;
--      call stg_makeStableName#(R1) returns to c1qb, args: 8, res: 8, upd: 8;
--
-- If we run sinking pass now (still before stack layout) we will get this:
--
--  c1q6:
--      I64[(young<c1q9> + 8)] = c1q9;
--      foreign call "ccall" arg hints:  []  result hints:  [] performMajorGC(...)
--                   returns to c1q9 args: ([]) ress: ([])ret_args: 8ret_off: 8;
--  c1q9:
--      I64[(young<c1qb> + 8)] = c1qb;
--      _s1pf::P64 = R1;         <------ _s1pf sunk past safe foreign call
--      R1 = _s1pc::P64;
--      call stg_makeStableName#(R1) returns to c1qb, args: 8, res: 8, upd: 8;
--
-- Notice that _s1pf was sunk past a foreign call. When we run stack layout
-- safe call to performMajorGC will be turned into:
--
--  c1q6:
--      _s1pc::P64 = P64[Sp + 8];
--      I64[Sp - 8] = c1q9;
--      Sp = Sp - 8;
--      I64[I64[CurrentTSO + 24] + 16] = Sp;
--      P64[CurrentNursery + 8] = Hp + 8;
--      (_u1qI::I64) = call "ccall" arg hints:  [PtrHint,]
--                           result hints:  [PtrHint] suspendThread(BaseReg, 0);
--      call "ccall" arg hints:  []  result hints:  [] performMajorGC();
--      (_u1qJ::I64) = call "ccall" arg hints:  [PtrHint]
--                           result hints:  [PtrHint] resumeThread(_u1qI::I64);
--      BaseReg = _u1qJ::I64;
--      _u1qK::P64 = CurrentTSO;
--      _u1qL::P64 = I64[_u1qK::P64 + 24];
--      Sp = I64[_u1qL::P64 + 16];
--      SpLim = _u1qL::P64 + 192;
--      HpAlloc = 0;
--      Hp = I64[CurrentNursery + 8] - 8;
--      HpLim = I64[CurrentNursery] + (%MO_SS_Conv_W32_W64(I32[CurrentNursery + 48]) * 4096 - 1);
--      call (I64[Sp])() returns to c1q9, args: 8, res: 8, upd: 8;
--  c1q9:
--      I64[(young<c1qb> + 8)] = c1qb;
--      _s1pf::P64 = R1;         <------ INCORRECT!
--      R1 = _s1pc::P64;
--      call stg_makeStableName#(R1) returns to c1qb, args: 8, res: 8, upd: 8;
--
-- Notice that c1q6 now ends with a call. Sinking _s1pf::P64 = R1 past that
-- call is clearly incorrect. This is what would happen if we assumed that
-- safe foreign call has the same semantics as unsafe foreign call. To prevent
-- this we need to treat safe foreign call as if was normal call.
432 433 434 435

-----------------------------------
-- mapping Expr in CmmNode

436
mapForeignTarget :: (CmmExpr -> CmmExpr) -> ForeignTarget -> ForeignTarget
437 438 439 440
mapForeignTarget exp   (ForeignTarget e c) = ForeignTarget (exp e) c
mapForeignTarget _   m@(PrimTarget _)      = m

wrapRecExp :: (CmmExpr -> CmmExpr) -> CmmExpr -> CmmExpr
441 442 443
-- Take a transformer on expressions and apply it recursively.
-- (wrapRecExp f e) first recursively applies itself to sub-expressions of e
--                  then  uses f to rewrite the resulting expression
444 445 446 447 448
wrapRecExp f (CmmMachOp op es)    = f (CmmMachOp op $ map (wrapRecExp f) es)
wrapRecExp f (CmmLoad addr ty)    = f (CmmLoad (wrapRecExp f addr) ty)
wrapRecExp f e                    = f e

mapExp :: (CmmExpr -> CmmExpr) -> CmmNode e x -> CmmNode e x
Peter Wortmann's avatar
Peter Wortmann committed
449
mapExp _ f@(CmmEntry{})                          = f
450
mapExp _ m@(CmmComment _)                        = m
Peter Wortmann's avatar
Peter Wortmann committed
451
mapExp _ m@(CmmTick _)                           = m
452 453 454 455 456 457
mapExp f   (CmmAssign r e)                       = CmmAssign r (f e)
mapExp f   (CmmStore addr e)                     = CmmStore (f addr) (f e)
mapExp f   (CmmUnsafeForeignCall tgt fs as)      = CmmUnsafeForeignCall (mapForeignTarget f tgt) fs (map f as)
mapExp _ l@(CmmBranch _)                         = l
mapExp f   (CmmCondBranch e ti fi)               = CmmCondBranch (f e) ti fi
mapExp f   (CmmSwitch e tbl)                     = CmmSwitch (f e) tbl
458
mapExp f   n@CmmCall {cml_target=tgt}            = n{cml_target = f tgt}
459
mapExp f   (CmmForeignCall tgt fs as succ ret_args updfr intrbl) = CmmForeignCall (mapForeignTarget f tgt) fs (map f as) succ ret_args updfr intrbl
460 461 462 463

mapExpDeep :: (CmmExpr -> CmmExpr) -> CmmNode e x -> CmmNode e x
mapExpDeep f = mapExp $ wrapRecExp f

464 465 466 467 468 469 470 471
------------------------------------------------------------------------
-- mapping Expr in CmmNode, but not performing allocation if no changes

mapForeignTargetM :: (CmmExpr -> Maybe CmmExpr) -> ForeignTarget -> Maybe ForeignTarget
mapForeignTargetM f (ForeignTarget e c) = (\x -> ForeignTarget x c) `fmap` f e
mapForeignTargetM _ (PrimTarget _)      = Nothing

wrapRecExpM :: (CmmExpr -> Maybe CmmExpr) -> (CmmExpr -> Maybe CmmExpr)
472 473
-- (wrapRecExpM f e) first recursively applies itself to sub-expressions of e
--                   then  gives f a chance to rewrite the resulting expression
474 475 476 477 478
wrapRecExpM f n@(CmmMachOp op es)  = maybe (f n) (f . CmmMachOp op)    (mapListM (wrapRecExpM f) es)
wrapRecExpM f n@(CmmLoad addr ty)  = maybe (f n) (f . flip CmmLoad ty) (wrapRecExpM f addr)
wrapRecExpM f e                    = f e

mapExpM :: (CmmExpr -> Maybe CmmExpr) -> CmmNode e x -> Maybe (CmmNode e x)
Peter Wortmann's avatar
Peter Wortmann committed
479
mapExpM _ (CmmEntry{})              = Nothing
480
mapExpM _ (CmmComment _)            = Nothing
Peter Wortmann's avatar
Peter Wortmann committed
481
mapExpM _ (CmmTick _)               = Nothing
482 483 484 485 486
mapExpM f (CmmAssign r e)           = CmmAssign r `fmap` f e
mapExpM f (CmmStore addr e)         = (\[addr', e'] -> CmmStore addr' e') `fmap` mapListM f [addr, e]
mapExpM _ (CmmBranch _)             = Nothing
mapExpM f (CmmCondBranch e ti fi)   = (\x -> CmmCondBranch x ti fi) `fmap` f e
mapExpM f (CmmSwitch e tbl)         = (\x -> CmmSwitch x tbl)       `fmap` f e
487
mapExpM f (CmmCall tgt mb_id r o i s) = (\x -> CmmCall x mb_id r o i s) `fmap` f tgt
488 489 490 491
mapExpM f (CmmUnsafeForeignCall tgt fs as)
    = case mapForeignTargetM f tgt of
        Just tgt' -> Just (CmmUnsafeForeignCall tgt' fs (mapListJ f as))
        Nothing   -> (\xs -> CmmUnsafeForeignCall tgt fs xs) `fmap` mapListM f as
492
mapExpM f (CmmForeignCall tgt fs as succ ret_args updfr intrbl)
493
    = case mapForeignTargetM f tgt of
494 495
        Just tgt' -> Just (CmmForeignCall tgt' fs (mapListJ f as) succ ret_args updfr intrbl)
        Nothing   -> (\xs -> CmmForeignCall tgt fs xs succ ret_args updfr intrbl) `fmap` mapListM f as
496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514

-- share as much as possible
mapListM :: (a -> Maybe a) -> [a] -> Maybe [a]
mapListM f xs = let (b, r) = mapListT f xs
                in if b then Just r else Nothing

mapListJ :: (a -> Maybe a) -> [a] -> [a]
mapListJ f xs = snd (mapListT f xs)

mapListT :: (a -> Maybe a) -> [a] -> (Bool, [a])
mapListT f xs = foldr g (False, []) (zip3 (tails xs) xs (map f xs))
    where g (_,   y, Nothing) (True, ys)  = (True,  y:ys)
          g (_,   _, Just y)  (True, ys)  = (True,  y:ys)
          g (ys', _, Nothing) (False, _)  = (False, ys')
          g (_,   _, Just y)  (False, ys) = (True,  y:ys)

mapExpDeepM :: (CmmExpr -> Maybe CmmExpr) -> CmmNode e x -> Maybe (CmmNode e x)
mapExpDeepM f = mapExpM $ wrapRecExpM f

515 516 517
-----------------------------------
-- folding Expr in CmmNode

518
foldExpForeignTarget :: (CmmExpr -> z -> z) -> ForeignTarget -> z -> z
519 520 521 522
foldExpForeignTarget exp (ForeignTarget e _) z = exp e z
foldExpForeignTarget _   (PrimTarget _)      z = z

-- Take a folder on expressions and apply it recursively.
523 524
-- Specifically (wrapRecExpf f e z) deals with CmmMachOp and CmmLoad
-- itself, delegating all the other CmmExpr forms to 'f'.
525 526 527 528 529 530 531 532
wrapRecExpf :: (CmmExpr -> z -> z) -> CmmExpr -> z -> z
wrapRecExpf f e@(CmmMachOp _ es) z = foldr (wrapRecExpf f) (f e z) es
wrapRecExpf f e@(CmmLoad addr _) z = wrapRecExpf f addr (f e z)
wrapRecExpf f e                  z = f e z

foldExp :: (CmmExpr -> z -> z) -> CmmNode e x -> z -> z
foldExp _ (CmmEntry {}) z                         = z
foldExp _ (CmmComment {}) z                       = z
Peter Wortmann's avatar
Peter Wortmann committed
533
foldExp _ (CmmTick {}) z                          = z
534 535 536 537 538 539 540 541 542 543
foldExp f (CmmAssign _ e) z                       = f e z
foldExp f (CmmStore addr e) z                     = f addr $ f e z
foldExp f (CmmUnsafeForeignCall t _ as) z         = foldr f (foldExpForeignTarget f t z) as
foldExp _ (CmmBranch _) z                         = z
foldExp f (CmmCondBranch e _ _) z                 = f e z
foldExp f (CmmSwitch e _) z                       = f e z
foldExp f (CmmCall {cml_target=tgt}) z            = f tgt z
foldExp f (CmmForeignCall {tgt=tgt, args=args}) z = foldr f (foldExpForeignTarget f tgt z) args

foldExpDeep :: (CmmExpr -> z -> z) -> CmmNode e x -> z -> z
Simon Marlow's avatar
Simon Marlow committed
544
foldExpDeep f = foldExp (wrapRecExpf f)
Simon Marlow's avatar
Simon Marlow committed
545

Simon Marlow's avatar
Simon Marlow committed
546 547 548 549 550 551
-- -----------------------------------------------------------------------------

mapSuccessors :: (Label -> Label) -> CmmNode O C -> CmmNode O C
mapSuccessors f (CmmBranch bid)        = CmmBranch (f bid)
mapSuccessors f (CmmCondBranch p y n)  = CmmCondBranch p (f y) (f n)
mapSuccessors f (CmmSwitch e arms)     = CmmSwitch e (map (fmap f) arms)
552
mapSuccessors _ n = n
Simon Marlow's avatar
Simon Marlow committed
553

Peter Wortmann's avatar
Peter Wortmann committed
554 555 556 557
-- -----------------------------------------------------------------------------

-- | Tickish in Cmm context (annotations only)
type CmmTickish = Tickish ()
Peter Wortmann's avatar
Peter Wortmann committed
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674

-- | Tick scope identifier, allowing us to reason about what
-- annotations in a Cmm block should scope over. We especially take
-- care to allow optimisations to reorganise blocks without losing
-- tick association in the process.
data CmmTickScope
  = GlobalScope
    -- ^ The global scope is the "root" of the scope graph. Every
    -- scope is a sub-scope of the global scope. It doesn't make sense
    -- to add ticks to this scope. On the other hand, this means that
    -- setting this scope on a block means no ticks apply to it.

  | SubScope U.Unique CmmTickScope
    -- ^ Constructs a new sub-scope to an existing scope. This allows
    -- us to translate Core-style scoping rules (see @tickishScoped@)
    -- into the Cmm world. Suppose the following code:
    --
    --   tick<1> case ... of
    --             A -> tick<2> ...
    --             B -> tick<3> ...
    --
    -- We want the top-level tick annotation to apply to blocks
    -- generated for the A and B alternatives. We can achieve that by
    -- generating tick<1> into a block with scope a, while the code
    -- for alternatives A and B gets generated into sub-scopes a/b and
    -- a/c respectively.

  | CombinedScope CmmTickScope CmmTickScope
    -- ^ A combined scope scopes over everything that the two given
    -- scopes cover. It is therefore a sub-scope of either scope. This
    -- is required for optimisations. Consider common block elimination:
    --
    --   A -> tick<2> case ... of
    --     C -> [common]
    --   B -> tick<3> case ... of
    --     D -> [common]
    --
    -- We will generate code for the C and D alternatives, and figure
    -- out afterwards that it's actually common code. Scoping rules
    -- dictate that the resulting common block needs to be covered by
    -- both tick<2> and tick<3>, therefore we need to construct a
    -- scope that is a child to *both* scope. Now we can do that - if
    -- we assign the scopes a/c and b/d to the common-ed up blocks,
    -- the new block could have a combined tick scope a/c+b/d, which
    -- both tick<2> and tick<3> apply to.

-- Note [CmmTick scoping details]:
--
-- The scope of a @CmmTick@ is given by the @CmmEntry@ node of the
-- same block. Note that as a result of this, optimisations making
-- tick scopes more specific can *reduce* the amount of code a tick
-- scopes over. Fixing this would require a separate @CmmTickScope@
-- field for @CmmTick@. Right now we do not do this simply because I
-- couldn't find an example where it actually mattered -- multiple
-- blocks within the same scope generally jump to each other, which
-- prevents common block elimination from happening in the first
-- place. But this is no strong reason, so if Cmm optimisations become
-- more involved in future this might have to be revisited.

-- | Output all scope paths.
scopeToPaths :: CmmTickScope -> [[U.Unique]]
scopeToPaths GlobalScope           = [[]]
scopeToPaths (SubScope u s)        = map (u:) (scopeToPaths s)
scopeToPaths (CombinedScope s1 s2) = scopeToPaths s1 ++ scopeToPaths s2

-- | Returns the head uniques of the scopes. This is based on the
-- assumption that the @Unique@ of @SubScope@ identifies the
-- underlying super-scope. Used for efficient equality and comparison,
-- see below.
scopeUniques :: CmmTickScope -> [U.Unique]
scopeUniques GlobalScope           = []
scopeUniques (SubScope u _)        = [u]
scopeUniques (CombinedScope s1 s2) = scopeUniques s1 ++ scopeUniques s2

-- Equality and order is based on the head uniques defined above. We
-- take care to short-cut the (extremly) common cases.
instance Eq CmmTickScope where
  GlobalScope    == GlobalScope     = True
  GlobalScope    == _               = False
  _              == GlobalScope     = False
  (SubScope u _) == (SubScope u' _) = u == u'
  (SubScope _ _) == _               = False
  _              == (SubScope _ _)  = False
  scope          == scope'          = sort (scopeUniques scope) ==
                                      sort (scopeUniques scope')
instance Ord CmmTickScope where
  compare GlobalScope    GlobalScope     = EQ
  compare GlobalScope    _               = LT
  compare _              GlobalScope     = GT
  compare (SubScope u _) (SubScope u' _) = compare u u'
  compare scope scope'                   = compare (sort $ scopeUniques scope)
                                                   (sort $ scopeUniques scope')

instance Outputable CmmTickScope where
  ppr GlobalScope     = text "global"
  ppr (SubScope us s) = ppr s <> char '/' <> ppr us
  ppr combined        = parens $ hcat $ punctuate (char '+') $
                        map (hcat . punctuate (char '/') . map ppr . reverse) $
                        scopeToPaths combined

-- | Checks whether two tick scopes are sub-scopes of each other. True
-- if the two scopes are equal.
isTickSubScope :: CmmTickScope -> CmmTickScope -> Bool
isTickSubScope = cmp
  where cmp _              GlobalScope             = True
        cmp GlobalScope    _                       = False
        cmp (CombinedScope s1 s2) s'               = cmp s1 s' && cmp s2 s'
        cmp s              (CombinedScope s1' s2') = cmp s s1' || cmp s s2'
        cmp (SubScope u s) s'@(SubScope u' _)      = u == u' || cmp s s'

-- | Combine two tick scopes. This smart constructor will catch cases
-- where one tick scope is a sub-scope of the other already.
combineTickScopes :: CmmTickScope -> CmmTickScope -> CmmTickScope
combineTickScopes s1 s2
  | s1 `isTickSubScope` s2 = s2
  | s2 `isTickSubScope` s1 = s1
  | otherwise              = CombinedScope s1 s2