CmmOpt.hs 27.8 KB
Newer Older
1 2 3 4 5 6 7 8 9
-----------------------------------------------------------------------------
--
-- Cmm optimisation
--
-- (c) The University of Glasgow 2006
--
-----------------------------------------------------------------------------

module CmmOpt (
10 11 12 13 14
        cmmEliminateDeadBlocks,
        cmmMiniInline,
        cmmMachOpFold,
        cmmMachOpFoldM,
        cmmLoopifyForC,
15 16 17 18
 ) where

#include "HsVersions.h"

19
import OldCmm
20
import OldPprCmm
21
import CmmNode (wrapRecExp)
Simon Marlow's avatar
Simon Marlow committed
22
import CmmUtils
23
import StaticFlags
24 25

import UniqFM
Simon Marlow's avatar
Simon Marlow committed
26
import Unique
27
import Util
28
import FastTypes
29
import Outputable
30
import Platform
31
import BlockId
32

Simon Marlow's avatar
Simon Marlow committed
33
import Data.Bits
34
import Data.Maybe
35
import Data.List
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53

-- -----------------------------------------------------------------------------
-- Eliminates dead blocks

{-
We repeatedly expand the set of reachable blocks until we hit a
fixpoint, and then prune any blocks that were not in this set.  This is
actually a required optimization, as dead blocks can cause problems
for invariants in the linear register allocator (and possibly other
places.)
-}

-- Deep fold over statements could probably be abstracted out, but it
-- might not be worth the effort since OldCmm is moribund
cmmEliminateDeadBlocks :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmEliminateDeadBlocks [] = []
cmmEliminateDeadBlocks blocks@(BasicBlock base_id _:_) =
    let -- Calculate what's reachable from what block
54 55 56
        reachableMap = foldl' f emptyUFM blocks -- lazy in values
            where f m (BasicBlock block_id stmts) = addToUFM m block_id (reachableFrom stmts)
        reachableFrom stmts = foldl stmt [] stmts
57 58 59 60 61
            where
                stmt m CmmNop = m
                stmt m (CmmComment _) = m
                stmt m (CmmAssign _ e) = expr m e
                stmt m (CmmStore e1 e2) = expr (expr m e1) e2
62
                stmt m (CmmCall c _ as _) = f (actuals m as) c
63 64
                    where f m (CmmCallee e _) = expr m e
                          f m (CmmPrim _) = m
65 66 67
                stmt m (CmmBranch b) = b:m
                stmt m (CmmCondBranch e b) = b:(expr m e)
                stmt m (CmmSwitch e bs) = catMaybes bs ++ expr m e
68
                stmt m (CmmJump e _) = expr m e
dterei's avatar
dterei committed
69
                stmt m (CmmReturn) = m
70 71 72
                actuals m as = foldl' (\m h -> expr m (hintlessCmm h)) m as
                -- We have to do a deep fold into CmmExpr because
                -- there may be a BlockId in the CmmBlock literal.
73 74 75
                expr m (CmmLit l) = lit m l
                expr m (CmmLoad e _) = expr m e
                expr m (CmmReg _) = m
76
                expr m (CmmMachOp _ es) = foldl' expr m es
77 78
                expr m (CmmStackSlot _ _) = m
                expr m (CmmRegOff _ _) = m
79
                lit m (CmmBlock b) = b:m
80
                lit m _ = m
81 82 83 84 85 86 87 88
        -- go todo done
        reachable = go [base_id] (setEmpty :: BlockSet)
          where go []     m = m
                go (x:xs) m
                    | setMember x m = go xs m
                    | otherwise     = go (add ++ xs) (setInsert x m)
                        where add = fromMaybe (panic "cmmEliminateDeadBlocks: unknown block")
                                              (lookupUFM reachableMap x)
89
    in filter (\(BasicBlock block_id _) -> setMember block_id reachable) blocks
90 91 92 93

-- -----------------------------------------------------------------------------
-- The mini-inliner

Simon Marlow's avatar
Simon Marlow committed
94
{-
95 96 97 98
This pass inlines assignments to temporaries.  Temporaries that are
only used once are unconditionally inlined.  Temporaries that are used
two or more times are only inlined if they are assigned a literal.  It
works as follows:
Simon Marlow's avatar
Simon Marlow committed
99 100

  - count uses of each temporary
101
  - for each temporary:
Ian Lynagh's avatar
Ian Lynagh committed
102
        - attempt to push it forward to the statement that uses it
Simon Marlow's avatar
Simon Marlow committed
103
        - only push forward past assignments to other temporaries
Ian Lynagh's avatar
Ian Lynagh committed
104 105 106
          (assumes that temporaries are single-assignment)
        - if we reach the statement that uses it, inline the rhs
          and delete the original assignment.
Simon Marlow's avatar
Simon Marlow committed
107

108 109 110 111
[N.B. In the Quick C-- compiler, this optimization is achieved by a
 combination of two dataflow passes: forward substitution (peephole
 optimization) and dead-assignment elimination.  ---NR]

Simon Marlow's avatar
Simon Marlow committed
112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
Possible generalisations: here is an example from factorial

Fac_zdwfac_entry:
    cmG:
        _smi = R2;
        if (_smi != 0) goto cmK;
        R1 = R3;
        jump I64[Sp];
    cmK:
        _smn = _smi * R3;
        R2 = _smi + (-1);
        R3 = _smn;
        jump Fac_zdwfac_info;

We want to inline _smi and _smn.  To inline _smn:

   - we must be able to push forward past assignments to global regs.
     We can do this if the rhs of the assignment we are pushing
     forward doesn't refer to the global reg being assigned to; easy
     to test.

To inline _smi:

   - It is a trivial replacement, reg for reg, but it occurs more than
     once.
   - We can inline trivial assignments even if the temporary occurs
     more than once, as long as we don't eliminate the original assignment
     (this doesn't help much on its own).
   - We need to be able to propagate the assignment forward through jumps;
     if we did this, we would find that it can be inlined safely in all
     its occurrences.
-}

145 146 147 148
countUses :: UserOfLocalRegs a => a -> UniqFM Int
countUses a = foldRegsUsed (\m r -> addToUFM m r (count m r + 1)) emptyUFM a
  where count m r = lookupWithDefaultUFM m (0::Int) r

149 150
cmmMiniInline :: Platform -> [CmmBasicBlock] -> [CmmBasicBlock]
cmmMiniInline platform blocks = map do_inline blocks
151
  where do_inline (BasicBlock id stmts)
152
          = BasicBlock id (cmmMiniInlineStmts platform (countUses blocks) stmts)
153

154
cmmMiniInlineStmts :: Platform -> UniqFM Int -> [CmmStmt] -> [CmmStmt]
Ian Lynagh's avatar
Ian Lynagh committed
155
cmmMiniInlineStmts _        _    [] = []
156
cmmMiniInlineStmts platform uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
157 158
        -- not used: just discard this assignment
  | Nothing <- lookupUFM uses u
159
  = cmmMiniInlineStmts platform uses stmts
160

161 162 163 164 165 166 167 168 169 170 171
        -- used (literal): try to inline at all the use sites
  | Just n <- lookupUFM uses u, isLit expr
  =
     ncgDebugTrace ("nativeGen: inlining " ++ showSDoc (pprStmt platform stmt)) $
     case lookForInlineLit u expr stmts of
         (m, stmts')
             | n == m -> cmmMiniInlineStmts platform (delFromUFM uses u) stmts'
             | otherwise ->
                 stmt : cmmMiniInlineStmts platform (adjustUFM (\x -> x - m) uses u) stmts'

        -- used (foldable to literal): try to inline at all the use sites
172
  | Just n <- lookupUFM uses u,
173
    e@(CmmLit _) <- wrapRecExp foldExp expr
174
  =
175
     ncgDebugTrace ("nativeGen: inlining " ++ showSDoc (pprStmt platform stmt)) $
176
     case lookForInlineLit u e stmts of
177
         (m, stmts')
178
             | n == m -> cmmMiniInlineStmts platform (delFromUFM uses u) stmts'
179
             | otherwise ->
180
                 stmt : cmmMiniInlineStmts platform (adjustUFM (\x -> x - m) uses u) stmts'
181

182
        -- used once (non-literal): try to inline at the use site
183 184 185
  | Just 1 <- lookupUFM uses u,
    Just stmts' <- lookForInline u expr stmts
  = 
186 187
     ncgDebugTrace ("nativeGen: inlining " ++ showSDoc (pprStmt platform stmt)) $
     cmmMiniInlineStmts platform uses stmts'
188
 where
Ian Lynagh's avatar
Ian Lynagh committed
189
  foldExp (CmmMachOp op args) = cmmMachOpFold platform op args
190
  foldExp e = e
191

192 193 194 195
  ncgDebugTrace str x = if ncgDebugIsOn then trace str x else x

cmmMiniInlineStmts platform uses (stmt:stmts)
  = stmt : cmmMiniInlineStmts platform uses stmts
196

197 198 199 200
-- | Takes a register, a 'CmmLit' expression assigned to that
-- register, and a list of statements.  Inlines the expression at all
-- use sites of the register.  Returns the number of substituations
-- made and the, possibly modified, list of statements.
201 202 203
lookForInlineLit :: Unique -> CmmExpr -> [CmmStmt] -> (Int, [CmmStmt])
lookForInlineLit _ _ [] = (0, [])
lookForInlineLit u expr stmts@(stmt : rest)
204
  | Just n <- lookupUFM (countUses stmt) u
205
  = case lookForInlineLit u expr rest of
206 207 208
      (m, stmts) -> let z = n + m
                    in z `seq` (z, inlineStmt u expr stmt : stmts)

209 210
  | ok_to_skip
  = case lookForInlineLit u expr rest of
211 212 213 214
      (n, stmts) -> (n, stmt : stmts)

  | otherwise
  = (0, stmts)
215 216 217 218 219 220
  where
    -- We skip over assignments to registers, unless the register
    -- being assigned to is the one we're inlining.
    ok_to_skip = case stmt of
        CmmAssign (CmmLocal (LocalReg u' _)) _ | u' == u -> False
        _other -> True
221

Ian Lynagh's avatar
Ian Lynagh committed
222
lookForInline :: Unique -> CmmExpr -> [CmmStmt] -> Maybe [CmmStmt]
223 224 225
lookForInline u expr stmts = lookForInline' u expr regset stmts
    where regset = foldRegsUsed extendRegSet emptyRegSet expr

Ian Lynagh's avatar
Ian Lynagh committed
226 227
lookForInline' :: Unique -> CmmExpr -> RegSet -> [CmmStmt] -> Maybe [CmmStmt]
lookForInline' _ _    _      [] = panic "lookForInline' []"
228
lookForInline' u expr regset (stmt : rest)
229 230 231
  | Just 1 <- lookupUFM (countUses stmt) u, ok_to_inline
  = Just (inlineStmt u expr stmt : rest)

232
  | ok_to_skip
233
  = case lookForInline' u expr regset rest of
234 235 236 237 238
           Nothing    -> Nothing
           Just stmts -> Just (stmt:stmts)

  | otherwise 
  = Nothing
239

240
  where
Ian Lynagh's avatar
Ian Lynagh committed
241 242 243 244 245
        -- we don't inline into CmmCall if the expression refers to global
        -- registers.  This is a HACK to avoid global registers clashing with
        -- C argument-passing registers, really the back-end ought to be able
        -- to handle it properly, but currently neither PprC nor the NCG can
        -- do it.  See also CgForeignCall:load_args_into_temps.
246
    ok_to_inline = case stmt of
Ian Lynagh's avatar
Ian Lynagh committed
247 248
                     CmmCall{} -> hasNoGlobalRegs expr
                     _ -> True
249

250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
   -- Expressions aren't side-effecting.  Temporaries may or may not
   -- be single-assignment depending on the source (the old code
   -- generator creates single-assignment code, but hand-written Cmm
   -- and Cmm from the new code generator is not single-assignment.)
   -- So we do an extra check to make sure that the register being
   -- changed is not one we were relying on.  I don't know how much of a
   -- performance hit this is (we have to create a regset for every
   -- instruction.) -- EZY
    ok_to_skip = case stmt of
                 CmmNop -> True
                 CmmComment{} -> True
                 CmmAssign (CmmLocal r@(LocalReg u' _)) _rhs | u' /= u && not (r `elemRegSet` regset) -> True
                 CmmAssign g@(CmmGlobal _) _rhs -> not (g `regUsedIn` expr)
                 _other -> False

265

266 267 268
inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
269 270
inlineStmt u a (CmmCall target regs es ret)
   = CmmCall (infn target) regs es' ret
271
   where infn (CmmCallee fn cconv) = CmmCallee (inlineExpr u a fn) cconv
Ian Lynagh's avatar
Ian Lynagh committed
272 273
         infn (CmmPrim p) = CmmPrim p
         es' = [ (CmmHinted (inlineExpr u a e) hint) | (CmmHinted e hint) <- es ]
274 275
inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
276
inlineStmt u a (CmmJump e live) = CmmJump (inlineExpr u a e) live
Ian Lynagh's avatar
Ian Lynagh committed
277
inlineStmt _ _ other_stmt = other_stmt
278 279

inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
280
inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
281 282
  | u == u' = a
  | otherwise = e
283 284
inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
  | u == u' = CmmMachOp (MO_Add width) [a, CmmLit (CmmInt (fromIntegral off) width)]
285
  | otherwise = e
286 287
  where
    width = typeWidth rep
288 289
inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
Ian Lynagh's avatar
Ian Lynagh committed
290
inlineExpr _ _ other_expr = other_expr
291 292 293 294 295 296 297 298

-- -----------------------------------------------------------------------------
-- MachOp constant folder

-- Now, try to constant-fold the MachOps.  The arguments have already
-- been optimized and folded.

cmmMachOpFold
Ian Lynagh's avatar
Ian Lynagh committed
299 300
    :: Platform
    -> MachOp       -- The operation from an CmmMachOp
301
    -> [CmmExpr]    -- The optimized arguments
302 303
    -> CmmExpr

Ian Lynagh's avatar
Ian Lynagh committed
304
cmmMachOpFold platform op args = fromMaybe (CmmMachOp op args) (cmmMachOpFoldM platform op args)
305 306 307 308

-- Returns Nothing if no changes, useful for Hoopl, also reduces
-- allocation!
cmmMachOpFoldM
Ian Lynagh's avatar
Ian Lynagh committed
309 310
    :: Platform
    -> MachOp
311 312 313
    -> [CmmExpr]
    -> Maybe CmmExpr

Ian Lynagh's avatar
Ian Lynagh committed
314
cmmMachOpFoldM _ op [CmmLit (CmmInt x rep)]
315
  = Just $ case op of
Ian Lynagh's avatar
Ian Lynagh committed
316 317
      MO_S_Neg _ -> CmmLit (CmmInt (-x) rep)
      MO_Not _   -> CmmLit (CmmInt (complement x) rep)
318

319 320 321 322
        -- these are interesting: we must first narrow to the 
        -- "from" type, in order to truncate to the correct size.
        -- The final narrow/widen to the destination type
        -- is implicit in the CmmLit.
Ian Lynagh's avatar
Ian Lynagh committed
323 324 325
      MO_SF_Conv _from to -> CmmLit (CmmFloat (fromInteger x) to)
      MO_SS_Conv  from to -> CmmLit (CmmInt (narrowS from x) to)
      MO_UU_Conv  from to -> CmmLit (CmmInt (narrowU from x) to)
326

327
      _ -> panic "cmmMachOpFoldM: unknown unary op"
328 329 330


-- Eliminate conversion NOPs
Ian Lynagh's avatar
Ian Lynagh committed
331 332
cmmMachOpFoldM _ (MO_SS_Conv rep1 rep2) [x] | rep1 == rep2 = Just x
cmmMachOpFoldM _ (MO_UU_Conv rep1 rep2) [x] | rep1 == rep2 = Just x
333 334

-- Eliminate nested conversions where possible
Ian Lynagh's avatar
Ian Lynagh committed
335
cmmMachOpFoldM platform conv_outer [CmmMachOp conv_inner [x]]
336 337 338
  | Just (rep1,rep2,signed1) <- isIntConversion conv_inner,
    Just (_,   rep3,signed2) <- isIntConversion conv_outer
  = case () of
339 340 341 342 343 344
        -- widen then narrow to the same size is a nop
      _ | rep1 < rep2 && rep1 == rep3 -> Just x
        -- Widen then narrow to different size: collapse to single conversion
        -- but remember to use the signedness from the widening, just in case
        -- the final conversion is a widen.
        | rep1 < rep2 && rep2 > rep3 ->
Ian Lynagh's avatar
Ian Lynagh committed
345
            Just $ cmmMachOpFold platform (intconv signed1 rep1 rep3) [x]
346 347
        -- Nested widenings: collapse if the signedness is the same
        | rep1 < rep2 && rep2 < rep3 && signed1 == signed2 ->
Ian Lynagh's avatar
Ian Lynagh committed
348
            Just $ cmmMachOpFold platform (intconv signed1 rep1 rep3) [x]
349 350
        -- Nested narrowings: collapse
        | rep1 > rep2 && rep2 > rep3 ->
Ian Lynagh's avatar
Ian Lynagh committed
351
            Just $ cmmMachOpFold platform (MO_UU_Conv rep1 rep3) [x]
352 353
        | otherwise ->
            Nothing
354
  where
355 356 357 358 359
        isIntConversion (MO_UU_Conv rep1 rep2) 
          = Just (rep1,rep2,False)
        isIntConversion (MO_SS_Conv rep1 rep2)
          = Just (rep1,rep2,True)
        isIntConversion _ = Nothing
360

361 362
        intconv True  = MO_SS_Conv
        intconv False = MO_UU_Conv
363 364 365 366 367

-- ToDo: a narrow of a load can be collapsed into a narrow load, right?
-- but what if the architecture only supports word-sized loads, should
-- we do the transformation anyway?

Ian Lynagh's avatar
Ian Lynagh committed
368
cmmMachOpFoldM _ mop [CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
369
  = case mop of
370 371
        -- for comparisons: don't forget to narrow the arguments before
        -- comparing, since they might be out of range.
Ian Lynagh's avatar
Ian Lynagh committed
372 373
        MO_Eq _   -> Just $ CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordWidth)
        MO_Ne _   -> Just $ CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordWidth)
374

Ian Lynagh's avatar
Ian Lynagh committed
375 376 377 378
        MO_U_Gt _ -> Just $ CmmLit (CmmInt (if x_u >  y_u then 1 else 0) wordWidth)
        MO_U_Ge _ -> Just $ CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordWidth)
        MO_U_Lt _ -> Just $ CmmLit (CmmInt (if x_u <  y_u then 1 else 0) wordWidth)
        MO_U_Le _ -> Just $ CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordWidth)
379

Ian Lynagh's avatar
Ian Lynagh committed
380 381 382 383
        MO_S_Gt _ -> Just $ CmmLit (CmmInt (if x_s >  y_s then 1 else 0) wordWidth)
        MO_S_Ge _ -> Just $ CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordWidth)
        MO_S_Lt _ -> Just $ CmmLit (CmmInt (if x_s <  y_s then 1 else 0) wordWidth)
        MO_S_Le _ -> Just $ CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordWidth)
384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400

        MO_Add r -> Just $ CmmLit (CmmInt (x + y) r)
        MO_Sub r -> Just $ CmmLit (CmmInt (x - y) r)
        MO_Mul r -> Just $ CmmLit (CmmInt (x * y) r)
        MO_U_Quot r | y /= 0 -> Just $ CmmLit (CmmInt (x_u `quot` y_u) r)
        MO_U_Rem  r | y /= 0 -> Just $ CmmLit (CmmInt (x_u `rem`  y_u) r)
        MO_S_Quot r | y /= 0 -> Just $ CmmLit (CmmInt (x `quot` y) r)
        MO_S_Rem  r | y /= 0 -> Just $ CmmLit (CmmInt (x `rem` y) r)

        MO_And   r -> Just $ CmmLit (CmmInt (x .&. y) r)
        MO_Or    r -> Just $ CmmLit (CmmInt (x .|. y) r)
        MO_Xor   r -> Just $ CmmLit (CmmInt (x `xor` y) r)

        MO_Shl   r -> Just $ CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
        MO_U_Shr r -> Just $ CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
        MO_S_Shr r -> Just $ CmmLit (CmmInt (x `shiftR` fromIntegral y) r)

Ian Lynagh's avatar
Ian Lynagh committed
401
        _          -> Nothing
402 403

   where
404 405 406 407 408
        x_u = narrowU xrep x
        y_u = narrowU xrep y
        x_s = narrowS xrep x
        y_s = narrowS xrep y

409 410 411 412 413 414

-- When possible, shift the constants to the right-hand side, so that we
-- can match for strength reductions.  Note that the code generator will
-- also assume that constants have been shifted to the right when
-- possible.

Ian Lynagh's avatar
Ian Lynagh committed
415
cmmMachOpFoldM platform op [x@(CmmLit _), y]
416
   | not (isLit y) && isCommutableMachOp op
Ian Lynagh's avatar
Ian Lynagh committed
417
   = Just (cmmMachOpFold platform op [y, x])
418 419 420 421 422 423 424 425 426 427 428 429 430 431

-- Turn (a+b)+c into a+(b+c) where possible.  Because literals are
-- moved to the right, it is more likely that we will find
-- opportunities for constant folding when the expression is
-- right-associated.
--
-- ToDo: this appears to introduce a quadratic behaviour due to the
-- nested cmmMachOpFold.  Can we fix this?
--
-- Why do we check isLit arg1?  If arg1 is a lit, it means that arg2
-- is also a lit (otherwise arg1 would be on the right).  If we
-- put arg1 on the left of the rearranged expression, we'll get into a
-- loop:  (x1+x2)+x3 => x1+(x2+x3)  => (x2+x3)+x1 => x2+(x3+x1) ...
--
432 433 434
-- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
-- PicBaseReg from the corresponding label (or label difference).
--
Ian Lynagh's avatar
Ian Lynagh committed
435
cmmMachOpFoldM platform mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
436
   | mop2 `associates_with` mop1
437
     && not (isLit arg1) && not (isPicReg arg1)
Ian Lynagh's avatar
Ian Lynagh committed
438
   = Just (cmmMachOpFold platform mop2 [arg1, cmmMachOpFold platform mop1 [arg2,arg3]])
439 440 441 442 443 444
   where
     MO_Add{} `associates_with` MO_Sub{} = True
     mop1 `associates_with` mop2 =
        mop1 == mop2 && isAssociativeMachOp mop1

-- special case: (a - b) + c  ==>  a + (c - b)
Ian Lynagh's avatar
Ian Lynagh committed
445
cmmMachOpFoldM platform mop1@(MO_Add{}) [CmmMachOp mop2@(MO_Sub{}) [arg1,arg2], arg3]
446
   | not (isLit arg1) && not (isPicReg arg1)
Ian Lynagh's avatar
Ian Lynagh committed
447
   = Just (cmmMachOpFold platform mop1 [arg1, cmmMachOpFold platform mop2 [arg3,arg2]])
448 449

-- Make a RegOff if we can
Ian Lynagh's avatar
Ian Lynagh committed
450
cmmMachOpFoldM _ (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
451
  = Just $ CmmRegOff reg (fromIntegral (narrowS rep n))
Ian Lynagh's avatar
Ian Lynagh committed
452
cmmMachOpFoldM _ (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
453
  = Just $ CmmRegOff reg (off + fromIntegral (narrowS rep n))
Ian Lynagh's avatar
Ian Lynagh committed
454
cmmMachOpFoldM _ (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
455
  = Just $ CmmRegOff reg (- fromIntegral (narrowS rep n))
Ian Lynagh's avatar
Ian Lynagh committed
456
cmmMachOpFoldM _ (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
457
  = Just $ CmmRegOff reg (off - fromIntegral (narrowS rep n))
458 459 460

-- Fold label(+/-)offset into a CmmLit where possible

Ian Lynagh's avatar
Ian Lynagh committed
461
cmmMachOpFoldM _ (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
462
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
Ian Lynagh's avatar
Ian Lynagh committed
463
cmmMachOpFoldM _ (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
464
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
Ian Lynagh's avatar
Ian Lynagh committed
465
cmmMachOpFoldM _ (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
466
  = Just $ CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))
467

468

469 470 471 472 473 474
-- Comparison of literal with widened operand: perform the comparison
-- at the smaller width, as long as the literal is within range.

-- We can't do the reverse trick, when the operand is narrowed:
-- narrowing throws away bits from the operand, there's no way to do
-- the same comparison at the larger size.
475

Ian Lynagh's avatar
Ian Lynagh committed
476 477 478 479
cmmMachOpFoldM platform cmp [CmmMachOp conv [x], CmmLit (CmmInt i _)]
  |     -- powerPC NCG has a TODO for I8/I16 comparisons, so don't try
    platformArch platform `elem` [ArchX86, ArchX86_64],
        -- if the operand is widened:
480 481 482 483 484 485
    Just (rep, signed, narrow_fn) <- maybe_conversion conv,
        -- and this is a comparison operation:
    Just narrow_cmp <- maybe_comparison cmp rep signed,
        -- and the literal fits in the smaller size:
    i == narrow_fn rep i
        -- then we can do the comparison at the smaller size
Ian Lynagh's avatar
Ian Lynagh committed
486
  = Just (cmmMachOpFold platform narrow_cmp [x, CmmLit (CmmInt i rep)])
487
 where
488
    maybe_conversion (MO_UU_Conv from to)
489 490
        | to > from
        = Just (from, False, narrowU)
491 492
    maybe_conversion (MO_SS_Conv from to)
        | to > from
493
        = Just (from, True, narrowS)
494

Simon Marlow's avatar
Simon Marlow committed
495 496
        -- don't attempt to apply this optimisation when the source
        -- is a float; see #1916
497
    maybe_conversion _ = Nothing
498

499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
        -- careful (#2080): if the original comparison was signed, but
        -- we were doing an unsigned widen, then we must do an
        -- unsigned comparison at the smaller size.
    maybe_comparison (MO_U_Gt _) rep _     = Just (MO_U_Gt rep)
    maybe_comparison (MO_U_Ge _) rep _     = Just (MO_U_Ge rep)
    maybe_comparison (MO_U_Lt _) rep _     = Just (MO_U_Lt rep)
    maybe_comparison (MO_U_Le _) rep _     = Just (MO_U_Le rep)
    maybe_comparison (MO_Eq   _) rep _     = Just (MO_Eq   rep)
    maybe_comparison (MO_S_Gt _) rep True  = Just (MO_S_Gt rep)
    maybe_comparison (MO_S_Ge _) rep True  = Just (MO_S_Ge rep)
    maybe_comparison (MO_S_Lt _) rep True  = Just (MO_S_Lt rep)
    maybe_comparison (MO_S_Le _) rep True  = Just (MO_S_Le rep)
    maybe_comparison (MO_S_Gt _) rep False = Just (MO_U_Gt rep)
    maybe_comparison (MO_S_Ge _) rep False = Just (MO_U_Ge rep)
    maybe_comparison (MO_S_Lt _) rep False = Just (MO_U_Lt rep)
    maybe_comparison (MO_S_Le _) rep False = Just (MO_U_Le rep)
    maybe_comparison _ _ _ = Nothing
516

517 518
-- We can often do something with constants of 0 and 1 ...

Ian Lynagh's avatar
Ian Lynagh committed
519
cmmMachOpFoldM _ mop [x, y@(CmmLit (CmmInt 0 _))]
520
  = case mop of
Ian Lynagh's avatar
Ian Lynagh committed
521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
        MO_Add   _ -> Just x
        MO_Sub   _ -> Just x
        MO_Mul   _ -> Just y
        MO_And   _ -> Just y
        MO_Or    _ -> Just x
        MO_Xor   _ -> Just x
        MO_Shl   _ -> Just x
        MO_S_Shr _ -> Just x
        MO_U_Shr _ -> Just x
        MO_Ne    _ | isComparisonExpr x -> Just x
        MO_Eq    _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_U_Gt  _ | isComparisonExpr x -> Just x
        MO_S_Gt  _ | isComparisonExpr x -> Just x
        MO_U_Lt  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_S_Lt  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_U_Ge  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_S_Ge  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_U_Le  _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_S_Le  _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        _ -> Nothing

cmmMachOpFoldM _ mop [x, (CmmLit (CmmInt 1 rep))]
543
  = case mop of
Ian Lynagh's avatar
Ian Lynagh committed
544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559
        MO_Mul    _ -> Just x
        MO_S_Quot _ -> Just x
        MO_U_Quot _ -> Just x
        MO_S_Rem  _ -> Just $ CmmLit (CmmInt 0 rep)
        MO_U_Rem  _ -> Just $ CmmLit (CmmInt 0 rep)
        MO_Ne    _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_Eq    _ | isComparisonExpr x -> Just x
        MO_U_Lt  _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_S_Lt  _ | Just x' <- maybeInvertCmmExpr x -> Just x'
        MO_U_Gt  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_S_Gt  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 0 wordWidth)
        MO_U_Le  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_S_Le  _ | isComparisonExpr x -> Just $ CmmLit (CmmInt 1 wordWidth)
        MO_U_Ge  _ | isComparisonExpr x -> Just x
        MO_S_Ge  _ | isComparisonExpr x -> Just x
        _ -> Nothing
560 561 562

-- Now look for multiplication/division by powers of 2 (integers).

Ian Lynagh's avatar
Ian Lynagh committed
563
cmmMachOpFoldM platform mop [x, (CmmLit (CmmInt n _))]
564
  = case mop of
565 566
        MO_Mul rep
           | Just p <- exactLog2 n ->
Ian Lynagh's avatar
Ian Lynagh committed
567
                 Just (cmmMachOpFold platform (MO_Shl rep) [x, CmmLit (CmmInt p rep)])
568 569
        MO_U_Quot rep
           | Just p <- exactLog2 n ->
Ian Lynagh's avatar
Ian Lynagh committed
570
                 Just (cmmMachOpFold platform (MO_U_Shr rep) [x, CmmLit (CmmInt p rep)])
571 572 573 574 575 576 577 578 579 580
        MO_S_Quot rep
           | Just p <- exactLog2 n, 
             CmmReg _ <- x ->   -- We duplicate x below, hence require
                                -- it is a reg.  FIXME: remove this restriction.
                -- shift right is not the same as quot, because it rounds
                -- to minus infinity, whereasq quot rounds toward zero.
                -- To fix this up, we add one less than the divisor to the
                -- dividend if it is a negative number.
                --
                -- to avoid a test/jump, we use the following sequence:
Ian Lynagh's avatar
Ian Lynagh committed
581
                --      x1 = x >> word_size-1  (all 1s if -ve, all 0s if +ve)
582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
                --      x2 = y & (divisor-1)
                --      result = (x+x2) >>= log2(divisor)
                -- this could be done a bit more simply using conditional moves,
                -- but we're processor independent here.
                --
                -- we optimise the divide by 2 case slightly, generating
                --      x1 = x >> word_size-1  (unsigned)
                --      return = (x + x1) >>= log2(divisor)
                let
                    bits = fromIntegral (widthInBits rep) - 1
                    shr = if p == 1 then MO_U_Shr rep else MO_S_Shr rep
                    x1 = CmmMachOp shr [x, CmmLit (CmmInt bits rep)]
                    x2 = if p == 1 then x1 else
                         CmmMachOp (MO_And rep) [x1, CmmLit (CmmInt (n-1) rep)]
                    x3 = CmmMachOp (MO_Add rep) [x, x2]
                in
Ian Lynagh's avatar
Ian Lynagh committed
598
                Just (cmmMachOpFold platform (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)])
Ian Lynagh's avatar
Ian Lynagh committed
599
        _ -> Nothing
600 601 602

-- Anything else is just too hard.

Ian Lynagh's avatar
Ian Lynagh committed
603
cmmMachOpFoldM _ _ _ = Nothing
604 605 606 607 608 609 610 611 612 613 614

-- -----------------------------------------------------------------------------
-- exactLog2

-- This algorithm for determining the $\log_2$ of exact powers of 2 comes
-- from GCC.  It requires bit manipulation primitives, and we use GHC
-- extensions.  Tough.
-- 
-- Used to be in MachInstrs --SDM.
-- ToDo: remove use of unboxery --SDM.

615 616 617 618 619
-- Unboxery removed in favor of FastInt; but is the function supposed to fail
-- on inputs >= 2147483648, or was that just an implementation artifact?
-- And is this speed-critical, or can we just use Integer operations
-- (including Data.Bits)?
--  --Isaac Dupree
620 621

exactLog2 :: Integer -> Maybe Integer
622 623
exactLog2 x_
  = if (x_ <= 0 || x_ >= 2147483648) then
624 625
       Nothing
    else
626 627
       case iUnbox (fromInteger x_) of { x ->
       if (x `bitAndFastInt` negateFastInt x) /=# x then
Ian Lynagh's avatar
Ian Lynagh committed
628
          Nothing
629
       else
Ian Lynagh's avatar
Ian Lynagh committed
630
          Just (toInteger (iBox (pow2 x)))
631 632
       }
  where
633 634
    pow2 x | x ==# _ILIT(1) = _ILIT(0)
           | otherwise = _ILIT(1) +# pow2 (x `shiftR_FastInt` _ILIT(1))
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


-- -----------------------------------------------------------------------------
-- Loopify for C

{-
 This is a simple pass that replaces tail-recursive functions like this:

   fac() {
     ...
     jump fac();
   }

 with this:

  fac() {
   L:
     ...
     goto L;
  }

  the latter generates better C code, because the C compiler treats it
  like a loop, and brings full loop optimisation to bear.

  In my measurements this makes little or no difference to anything
  except factorial, but what the hell.
-}

Simon Peyton Jones's avatar
Simon Peyton Jones committed
663
cmmLoopifyForC :: RawCmmDecl -> RawCmmDecl
664
cmmLoopifyForC p@(CmmProc Nothing _ _) = p  -- only if there's an info table, ignore case alts
Ian Lynagh's avatar
Ian Lynagh committed
665
cmmLoopifyForC (CmmProc (Just info@(Statics info_lbl _)) entry_lbl
666
                 (ListGraph blocks@(BasicBlock top_id _ : _))) =
667
--  pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
668
  CmmProc (Just info) entry_lbl (ListGraph blocks')
669
  where blocks' = [ BasicBlock id (map do_stmt stmts)
Ian Lynagh's avatar
Ian Lynagh committed
670
                  | BasicBlock id stmts <- blocks ]
671

672
        do_stmt (CmmJump (CmmLit (CmmLabel lbl)) _) | lbl == jump_lbl
Ian Lynagh's avatar
Ian Lynagh committed
673 674
                = CmmBranch top_id
        do_stmt stmt = stmt
675

Ian Lynagh's avatar
Ian Lynagh committed
676 677
        jump_lbl | tablesNextToCode = info_lbl
                 | otherwise        = entry_lbl
678 679 680 681 682 683

cmmLoopifyForC top = top

-- -----------------------------------------------------------------------------
-- Utils

Ian Lynagh's avatar
Ian Lynagh committed
684
isLit :: CmmExpr -> Bool
685 686 687 688 689
isLit (CmmLit _) = True
isLit _          = False

isComparisonExpr :: CmmExpr -> Bool
isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
Ian Lynagh's avatar
Ian Lynagh committed
690
isComparisonExpr _                  = False
691

Ian Lynagh's avatar
Ian Lynagh committed
692
isPicReg :: CmmExpr -> Bool
693
isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True
694
isPicReg _ = False
695